WO1999048018A1 - A scalable system for clustering of large databases - Google Patents

A scalable system for clustering of large databases Download PDF

Info

Publication number
WO1999048018A1
WO1999048018A1 PCT/US1999/005759 US9905759W WO9948018A1 WO 1999048018 A1 WO1999048018 A1 WO 1999048018A1 US 9905759 W US9905759 W US 9905759W WO 9948018 A1 WO9948018 A1 WO 9948018A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
database
clustering
model
cluster
Prior art date
Application number
PCT/US1999/005759
Other languages
French (fr)
Inventor
Usama Fayyad
Paul S. Bradley
Cory Reina
Original Assignee
Microsoft Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Microsoft Corporation filed Critical Microsoft Corporation
Priority to EP99913907A priority Critical patent/EP1062590B1/en
Priority to DE69938339T priority patent/DE69938339T2/en
Publication of WO1999048018A1 publication Critical patent/WO1999048018A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35Clustering; Classification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2321Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2216/00Indexing scheme relating to additional aspects of information retrieval not explicitly covered by G06F16/00 and subgroups
    • G06F2216/03Data mining
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y10TECHNICAL SUBJECTS COVERED BY FORMER USPC
    • Y10STECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y10S707/00Data processing: database and file management or data structures
    • Y10S707/99941Database schema or data structure
    • Y10S707/99942Manipulating data structure, e.g. compression, compaction, compilation

Definitions

  • the present invention concerns database analysis and more particularly concerns an apparatus and method for clustering of data into groups that capture important regularities and characteristics of the data.
  • DBMS database management systems
  • Clustering An important class of problems in the areas of decision support and reporting are clustering (segmentation) problems where one is interested in finding groupings (clusters) in the data. Clustering has been studied for decades in statistics, pattern recognition, machine learning, and many other fields of science and engineering. However, implementations and applications have historically been limited to small data sets with a small number of dimensions.
  • Each cluster includes records that are more similar to members of the same cluster than they are similar to rest of the data. For example, in a marketing application, a company may want to decide who to target for an ad campaign based on historical data about a set of customers and how they responded to previous campaigns. Other examples of such problems include: fraud detection, credit approval, diagnosis of system problems, diagnosis of manufacturing problems, recognition of event signatures, etc.
  • Employing analysts (statisticians) to build cluster models is expensive, and often not effective for large problems (large data sets with large numbers of fields). Even trained scientists can fail in the quest for reliable clusters when the problem is high-dimensional (i.e. the data has many fields, say more than 20).
  • Clustering is a necessary step in the mining of large databases as it represents a means for finding segments of the data that need to be modeled separately. This is an especially important consideration for large databases where a global model of the entire data typically makes no sense as data represents multiple populations that need to be modeled separately. Random sampling cannot help in deciding what the clusters are. Finally, clustering is an essential step if one needs to perform density estimation over the database (i.e. model the probability distribution governing the data source). Applications of clustering are numerous and include the following broad areas: data mining, data analysis in general, data visualization, sampling, indexing, prediction, and compression.
  • Clustering is an important area of application for a variety of fields including data mining, statistical data analysis, compression, and vector quantization.
  • Clustering has been formulated in various ways.
  • the fundamental clustering problem is that of grouping together (clustering) data items that are similar to each other.
  • the most general approach to clustering is to view it as a density estimation problem.
  • the data is assumed to arrive from a mixture model and the mixing labels (cluster identifiers) are hidden.
  • K-Means a data item belongs to a single cluster, while in EM each data item is assumed to belong to every cluster but with a different probability. This of course affects the update step (3) of the algorithm.
  • each cluster is updated based strictly on its membership.
  • each cluster is updated by the entire data set according to the relative probability of membership.
  • the invention represents a methodology for scaling clustering algorithms to large databases.
  • the invention enables effective and accurate clustering in one or less scans of a database.
  • Use of the invention results in significantly better performance than prior art schemes that are based on random sampling. These results are achieved with significantly less memory requirement and acceptable accuracy in terms of approaching the true solution than if one had run the clustering algorithm on the entire database.
  • Known methods can only address small databases (ones that fit in memory) or resort to sampling only a fraction of the data.
  • the disclosed invention is based on the concept of retaining in memory only the data points that need to be present in memory. The majority of the data points are summarized into a condensed representation that represents their sufficient statistics. By analyzing a mixture of sufficient statistics and actual data points, significantly better clustering results than random sampling methods are achieved and with similar lower memory requirements.
  • the invention can typically terminate well before scanning all the data in the database, hence gaining a major advantage over other scalable clustering methods that require at a minimum a full data scan.
  • the invention concerns a framework that supports a wide class of clustering algorithms.
  • the K-means algorithm as an example clustering algorithm that represents one specific embodiment of this framework.
  • the framework is intended to support a variety of algorithms that can be characterized by iteratively scanning data and updating models.
  • K-Means since it is well-known and established clustering method originally known as Forgy's method and has been used extensively in pattern recognition. It is a standard technique for clustering, used in a wide array of applications and even as way to initialize the more expensive EM clustering algorithm.
  • a clustering session may take days or weeks, and it is often desirable to update existing models as data arrives.
  • a list of desirable data mining characteristics follows: The invention satisfies all these:
  • Clustering should run within one scan (or less) of the database if possible: a single data scan is considered costly, early termination if appropriate is highly desirable.
  • the technique embodied in the invention relies on the observation that clustering techniques do not need to rescan all the data items as it is originally defined and as implemented in popular literature and statistical libraries and analysis packages.
  • the disclosed process may be viewed as an intelligent sampling scheme that employs some theoretically justified criteria for deciding which data can be summarized and represented by a significantly compressed set of sufficient statistics, and which data items must be carried in computer memory, and hence occupying a valuable resource.
  • DS discard set
  • CS compression set
  • RS retained set
  • the DS is summarized in a single set of sufficient statistics.
  • the compression set CS is summarized by multiple sufficient statistics representing subclusters of the CS data set.
  • the invention operates by obaining a next available (possibly random) sample from a database to fill free space in buffer. A current model of the clustering is then updated over the contents of the buffer. Elements of the new sample are identified to determine whether they need to be retained in the buffer (retained set RS); they can be discarded with updates to the sufficient statistics (discard set DS); or they can be reduced via compression and summarized as sufficient statistics (compression set CS). Once this has been done a determination is made to see if a stopping criteria is satisfied. If so terminate clustering, if not then sample more data.
  • the exemplary embodiment satisfies the above-mentioned important issues faced during data mining.
  • a clustering session on a large database can take days or even weeks. It is often desirable to update the clustering models as the data arrives and is stored. It is important in this data mining environment to be able to cluster in one scan (or less) of the database. A single scan is considered costly and clustering termination before one complete scan is highly desirable.
  • An exemplary embodiment of the invention includes a model optimizer. A multiple number of different clustering models are simultaneously generated in one or less scans of the database. The clustering analysis stops when one of the models reaches a stopping criteria. Alternately, the clustering can continue until all of the multiple models are complete as judged by the stopping criteria.
  • Figure 1 is a schematic illustration of a computer system for use in practicing the present invention
  • Figure 2 is schematic depiction of a clustering process of the invention wherein multiple clustering models are obtained during a single scan of the a database;
  • FIG. 3 A and 3B are schematic depictions showing software components of the invention.
  • Figure 4 is a flow diagram of an exemplary embodiment of the invention
  • Figure 5 is a one-dimensional plot of data distribution of three clusters of data
  • Figure 6 A - 6D are illustrations of data structures utilized for storing data in accordance with the exemplary embodiment of the invention
  • Figure 7 is a flow diagram of an exemplary embodiment of an extended K- means analysis of data
  • Figure 8 is a data structure for use in determining multiple data models through practice of the exemplary embodiment of the present invention.
  • Figure 9 is a plot of data points in two dimensions showing a clustering of data;
  • Figures 10 - 14 are user interface screens indicating various aspects of the clustering process.
  • a large database 10 for which the present invention has particular utility contains many records stored on multiple, possibly distributed storage devices. Each record has many attributes or fields which for a representative database might include age, income, number of children, number of cars owned etc.
  • a goal of the invention is to characterize clusters of data in the database 10. This task is straightforward for small databases (all data fits in the memory of a computer for example) having records that have a small number of fields or attributes. The task becomes very difficult, however, for large databases having huge numbers of records with a high dimension of attributes.
  • FIG 4 is a flow chart of the process steps performed during a scalable clustering analysis of data in accordance with the present invention.
  • An initialization step 100 includes a step of initializing a number of data structures shown in Figure 6A - 6D and choosing a cluster number K for characterizing the data.
  • a next step 1 10 is to sample a portion of the data in the database 10 from a storage medium to bring that portion of data within a random access memory (into RAM for example, although other forms of random access memory are contemplated) of the computer 20 shown in Figure 1.
  • the data has a large number of fields so that instead of a single dimension analysis, the invention characterizes a large number of vectors where the dimension of the vector is the number of attributes of the data records in the database.
  • the data structure 180 for this data is shown in Figure 6C to include a number r of records having a potentially large number of attributes.
  • the gathering of data can be performed in either a sequential scan that uses only a forward pointer to sequentially traverse the data or an indexed scan that provides a random sampling of data from the database.
  • the index scan it is a requirement that data not be accessed multiple times. This can be accomplished by marking data tuples to avoid duplicates, or a random index generator that does not repeat. In particular, it is most preferable that the first iteration of sampling data be done randomly. If it is known the data is random within the database then sequential scanning is acceptable. If it is not known that the data is randomly distributed, then random sampling is needed to avoid an inaccurate representative of the database.
  • a processor unit 21 of the computer 20 next executes 120 a clustering procedure using the data brought into memory in the step 110 as well as compressed data in the CS and DS data structures.
  • the processor 21 assigns data contained within the portion of data brought into memory to a cluster and determines a set of parameters for each attribute of the data assigned to a given cluster.
  • a data structure for the results or output model of the analysis is a model of the clustering of data and is depicted in Figure 6D. This model includes K records, one for each cluster.
  • each record has three required components: 1 ) a vector 'Sum' representing the sum for each of the attributes or dimensions of the vector data records in a given cluster, 2) a vector 'Sumsq' representing the sum of the attributes squared and 3) an integer 'M' counting the number of data records contained in or belonging to the corresponding cluster.
  • a vector 'Sum' representing the sum for each of the attributes or dimensions of the vector data records in a given cluster
  • a vector 'Sumsq' representing the sum of the attributes squared
  • a next step 130 in the Figure 4 flowchart summarizes at least some of the data used in the present iteration to characterize the K clusters.
  • This summarization is contained in the data structures DS, CS of Figures 6A and 6B.
  • the summarization takes up significantly less storage in a computer memory 25 than the vector data structure (Figure 6D) needed to store individual records. Storing a summarization of the data in the data structures of Figures 6 A and 6B frees up more memory allowing additional data to be sampled from the database 10.
  • the processor 21 determines 140 whether a stopping criteria has been reached.
  • One stopping criteria that is used is whether the analysis has produced a good enough model ( Figure 6D) by a standard that is described below.
  • a second stopping criteria has been reached if all the data in the database 10 has been used in the analysis.
  • One important aspect of the invention is the fact that instead of stopping the analysis, the analysis can be suspended. Data in the data structures of Figure 6A - 6D can be saved (either in memory or to disk) and the scalable clustering analysis can then be resumed later. This allows the database to be updated and the analysis resumed to update the clustering statistics without starting from the beginning. It also allows another process to take control of the processor 21 without losing the state of the clustering analysis.
  • the suspension could also be initiated in response to a user request that the analysis be suspended by means of a user actuated control on an interface presented to the user on a monitor 47 while the Clustering analysis is being performed.
  • FIGs 3 A and 3B depict an operating environment of the invention. Tjis is shown also in NEWFIG. Data in the database 10 is accessed through a database management system that sends data to a data mining engine 12. The data mining engine 12 processes requests from an application 14 and responds by sending back model summaries to the application. The software components of the data mining engine 12 are depicted in more detail in Figure 3 A.
  • K-Means Clustering There is a large literature of known data clustering techniques. One known technique is the so called K-means clustering process which is summarized in Duda- Hart (Pattern Classification and Scene Analysis) 1973 John Wiley & Sons, Inc., New York. An exemplary embodiment of the present scalable clustering analysis is described by reference to the K-Means Clustering process.
  • the K-Means algorithm takes as input: the number of clusters K, a set of K initial estimates of the cluster means, and the data set to be clustered.
  • centroids define the parameters of the model.
  • K-means evaluation starts with a random choice of cluster centroids or means that are randomly placed within the extent of the data on the x axis. Call these M l, M2, and M3 in Figure 9.
  • Each cluster is represented by its mean and each data item is assigned membership in the cluster having the nearest mean.
  • Distance is the Euclidean distance (or L2 norm): for a data point (d-dimensional vector) x and mean ⁇ , is given by:
  • K-Means is only defined over numeric (continuous-valued) data since the ability to compute the mean is a requirement.
  • a discrete version of K-Means exists and is sometimes referred to as
  • the K-Means algorithm finds a locally optimal solution to the problem of minimizing the sum of the L2 distance between each data point and its nearest cluster center (usually termed "distortion") .
  • K-means clustering centroids For a database that fits in memory all data within the database can be used to calculate the K-means clustering centroids.
  • the output from such a prior art process will be the K centroids and the number of data points that fall within each of the K clusters.
  • the data from the database 10 is brought into a memory 22 ( Figure 1) and once the K-means calculation is performed it is assigned to one of three data sets.
  • a retained data set (RS) is kept in memory 22 for further use in performing the K-means analysis.
  • a so called discard data set (DS) and a compressed data set (CS) are summarized in the form of sufficient statistics that are retained in memory. For a random sampling of data the more data that is gathered the better the clustering calculation becomes. For a process in accordance with the present invention the estimates do get better as more data is sampled, but the estimates are accurate even though memory requirements limit sampling to a low percentage of the entire data base.
  • Figures 6A - 6D summarize the data structures used to perform the scalable K- means analysis.
  • the output from the analysis is stored in a data structure designated MODEL ( Figure 6D) which includes an array 152 of pointers, each pointer points to a first vector 154 of n elements (floats) 'Sum', a second vector 156 of n elements (floats) 'SumSq', and a scalar 158 designated M.
  • the number n corresponds to the number of attributes of the database records that are being clustered.
  • Figure 6A depicts a data structure designated DS including an array 160 of pointers, each of which identifies a vector 162 of n elements (floats) 'Sum' , a vector 164 of n elements (floats) 'SumSq', and a scalar 166 designated 'M'.
  • a further data structure designated CS is an array of c pointers 170, where each pointer points to a vector 172 of n elements (floats) 'Sum', a vector 174 of n elements (floats) 'SumSq', and a scalar 176 designated as M.
  • An additional data structure designated RS ( Figure 6C) is an array 180 of r elements where each element points a vector of n elements (floats) representing a singleton data point of a type designated SDATA. As data is read in from the database it is stored in the set RS and this data is hot associated with any of the K clusters.
  • multiple models are simultaneously generated. This structure is depicted in Figure 8 wherein multiple models are pointers to arrays of models such as the data structures shown in Figure 6D.
  • Table 1 below is a list often SDATA vectors which constitute sample data from a database 10 and are stored as individual vectors in the data structure RS.
  • An important concept of the present invention is the summarization or compression of data points of type SDATA contained in the dataset RS (TABLE 1) sampled from the database into the two data structures DS, CS to allow more sampling of data from the database.
  • the scalable K-means analysis calls an 'extended K-means' procedure 120 that utilizes the compressed data as well as the remaining data samples contained in the dataset RS.
  • Two conceptual data structures help explain a first embodiment of the method of updating the set DS.
  • This first embodiment is referred to as the Worst Case Analysis method.
  • a list structure designated LOWER is a vector of k elements (one for each cluster) where each element points to a vector of n elements (floats) holding the lower bounds for each attribute of the CI on the mean of the corresponding cluster.
  • LOWER(3).LowVec(2) is the value of the lower bound on the CI for the third cluster along dimension 2.
  • a second structure designated UPPER is a vector of K elements (one for each cluster) where each element points to a vector of n elements (floats) holding the upper bounds for the CI on the parameters of the model (mean or centroid in case of K-means) of the corresponding cluster.
  • Singleton Points (Elements of RS) not changing cluster assignment when the K cluster centers are perturbed, within their respective confidence intervals in a worst-case fashion, can be summarized by adding them to the set DS and removing them from RS.
  • Appendix A is a summarization of the Worst Case Analysis that defines LOWER and UPPER as well as the process of updating of the set DS using the Worst Case Analysis.
  • a second embodiment of the process of updating the dataset DS is referred to as the Threshold Analysis.
  • a data structure is used that helps sort the singleton elements of RS (of type SDATA) by their Mahalanobis distance ( See Duda and Hart, Pattern Classification and Scene Analysis referenced above) to a given cluster center.
  • a structure RSDist is a list of r elements (r is the number of singleton data elements in the set RS) where each element in RSDist points to 2 objects: 1 ) float called "MahalDist" which holds the Mahalanobis distance of the corresponding element in RS to the nearest cluster center and 2) an integer indicating the cluster whose center is nearest to the given point in RS, called "Oust Assign".
  • Appendix C summarizes the calculation of the Mahalanobis distances.
  • Mahanalobis Distance is used in case of K-means as an example. In general, we require the ability to measure the distance of a point to a cluster. This distance can be obtained from any model as the probability that a data item belongs to the cluster. Low probability data items are FAR, while high probability items are CLOSE according to this distance. Our framework for scaling clustering in general requires ability to extract this notion of "distance”. This is always possible, regardless of model being used.
  • the Mahalanobis distance in case when clusters are represented by multivariate Gaussians is directly proportional to probability assigned to a data item by the Gaussian in that cluster.
  • a third embodiment for updating the dataset DS is based upon the distance of a data point from the then current parameters characterizing the cluster (e.g. the cluster mean in K-means). If this distance for a point rs contained in RS is less than a specified distance, then the point is removed from RS and added to DS by updating the sufficient statistics of DS to include the rs point.
  • the cluster mean in K-means e.g. the cluster mean in K-means
  • Figure 5 depicts a clustering in one dimensional, for example, income data from the table 1 list of data.
  • Three clusters K l, K2, K3 are made up of much larger numbers of data vectors SDATA.
  • the data structure DS that summarizes data for the cluster designated K 1 is centered within the generally Gaussian shaped curve depicting that cluster. Regardless of the technique used, certain data safely 'belongs' to the
  • Subclusters in dataset CS After the compression of data into the DS data structure, there is still a fairly large amount of data (SDATA) contained to the left and the right of the centroid of Kl in Figure 5 that neither the worst case analysis (Appendix A) nor the threshold analysis (Appendix B) identifies for compression into the set DS. These point fall within the 'compression region' for cluster 1.
  • the present invention also summarizes at least some of this data in the form of a dataset designated CS.
  • An exemplary process for determining the CS dataset is summarized in the psuedocode of Appendix C. Briefly, a dataset RSNew is made up of the dataset RS after removing the set DS from the original set RS.
  • the process of determining the new CS data set begins by merging and removing any singleton points from RSNew into CS which can safely be merged without violating a specified "density" criteria. For each data point in RSNew, the appendix C procedure finds the CS subcluster which is closest to it. If the data point can be merged into CS without violating the specified "density criteria" of the CS subcluster, then the data point is merged into that particular CS subcluster and removed from RSNew. If it doesn't satisfy the criteria, then the data point is left in RSNew.
  • the process of determining the data set CS then continues by finding a set of "dense" subclusters within the set RSNew. This is done by performing a traditional K- means analysis on the data in the set RSNew using a cluster number K' (Kprime) greater than K, the number of clusters used in performing the scalable K-means analysis.
  • K' Kprime
  • K cluster number
  • the set of sufficient statistics (Sum, Sumsq, and M) for the K' subclusters found by this procedure are appended to the current set of sufficient statistics in the dataset CS.
  • CS is augmented by K' elements to produce a number c of subclusters. This augmented list is then filtered and elements are merged (if possible), reducing the size of the list.
  • clustering algorithm being used in the main scalable framework (i.e. the clustering algorithm being scaled to large DB ), the clustering algorithm used to determine CS can be different. K-means is disclosed for secondary clustering (determining CS), but other clustering algorithms can be
  • MinPoints in Appendix C the data from this analysis is not clustered and the data is instead kept in the set RSNew. Furthermore only dense clusters are kept in CS. For each subcluster remaining after the threshold value of points criteria has been checked, if the maxiumum measure of spread computed from (SUMSQ) along any of the n dimensions (attributes) of the candidate subcluster is less than a threshold (StdTol in appendix C) the data from that subcluster is also left in RSNew and not summarized in CS. These two criteria remove the elements of CS corresponding to subclusters which are either too small (in terms of number of points) or too 'spread out'
  • a final step is the combining of subclusters using hierarchical agglomerative clustering.
  • An attempt is made to merge two elements of the dataset CS. If the larger, merged subcluster still satisfies the 'spread' criteria discussed above, these two subclusters are removed from the set CS, and the larger subcluster representing these two smaller subclusters is added to the dataset CS. This process continues until the merging process produces no larger subclusters that still satisfy the "dense" criteria.
  • the data structure CS contains c elements that are then used in the K-means analysis.
  • An extended CLUSTERING procedure 120 includes looping constructs for updating the current model that are summarized in the flow chart of Figure 7. Certain utility functions are needed to perform this extended CLUSTERING procedure.
  • a function ModelCopy( ModelNew , ModelOrig ) copies the model ModelOrig into
  • Zero( Model ) takes the data structure for the model in Figure 6D and sets all elements to 0.0.
  • a function Distance2Norm( point 1, point2) measures the distance between two vectors point 1 and point2.
  • the extended K-means procedure of Figure 7 begins with a step 200 of copying the existing model into an old model data structure. The process next
  • the 16 determines 202 the length of the RS, DS, and CS data structures of Figures 6A - 6C and returns the values k, c, and r.
  • the data structure NewModel is then zeroed or initialized 204.
  • the process updates the NewModel until a test 206 indicates a stopping criteria has been met. If the stopping criteria has not been met, the process saves 208 the new model in the old model data structure and branches 210 to zero a next subsequent new model.
  • the test 206 is similar to the test 140 described in conjunction with the scalable K-means process overview.
  • That model is updated in a loop 220 that updates the model for each of the r vectors in the dataset RS.
  • the loop gathers data 222 a point at a time from the set RS and determines 224 what cluster to assign that data point to. This is done by finding the distance from the data point to the mean of each of the then existing means of the old model.
  • Figure 6D for particular example of sclaing K-means, hence EXTENDED K_MEANS algorithm
  • the model includes the Sum for a given cluster K and therfore the mean or centroid of each dimension is given by this value divided by the scalar M for the cluster.
  • the New model Sum and SumSq component for that closest cluster is updated by adding the data point to the vector Cluster(closest).sum and then squaring the components and adding them to the Cluster(closest). SumSq vector.
  • the scalar M for a cluster is incremented by one for each point added to that cluster.
  • the procedure updates the model based on the compressed statistics stored in the c subclusters found in the data structure CS.
  • Table 3 indicates the contents of the model ( Figure 6D) after the RS portion of the extended K-means process on data points of Table 1.
  • the cluster mean for the second cluster K2 is at an income of 45.5k dollars. Data falling between the two means of 19.75 and 45.5 typically will not be classed in either the DS or the CS dataset. It is retained in RS and used on the next iteration to perform the clustering.
  • the CS and DS structures contain sufficient statistics, and the extended clustering procedure of Figure 7 must take into account this data in determining the new model when the procedure is called at the step 120 of the scalable K-means analysis.
  • the Figure 7 procedure executes a loop 230 over each of the subclusters c in CS and determines which of the K clusters in the Model ( Figure 6D) that subcluster is closest to. Assume subcluster p is closest to cluster q.
  • the sufficient statistics of cluster q are updated by adding the contents subcluster(p).sum to cluster(q).sum and the statistics subcluster(p).sumsq to cluster(q).sumsq. Additionally, the value of subcluster(p).M for the subcluster is added to the to the value cluster(q).M.
  • the extended CLSUTERING procedure updates the NewModel for the clusters summarized in DS.
  • the step 240 merely loops over the clusters in DS and adds their sufficient statistics to the new model of figure 6D.
  • a variable CenterDist is set to zero and for each of the clusters K, a distance between the centroid of the old model and the centroid of the new model is determined and added to the CenterDist variable. Once all K distances have been calculated, and added together the CenterDist value is divided by the number of clusters K and compared to a value 'StopTol' which is used as a measure of how stable the model has become. If the value of CenterDist is smaller than the value 'StopTol' then the procedure returns, otherwise the procedure branches back to recalculate the model using the same data in RS, CS, and DS but with a different "old model".
  • a first method terminates the analysis if the difference between the K- means, measured in some norm, over a given number of data samples is below a given tolerance.
  • a second method terminates if the difference in an "energy" function (measure of distortion)
  • a fourth stopping criteria is actually a suspension of the scalable K-means analysis rather than stopping.
  • PrevModelDiff in the case that the first stopping criteria is chosen , (see appendix D) from these z models or by computing PrevEnergyDiff (in the case that the second stopping criteria is chosen, see appendix D).
  • the computer 20 includes a monitor 47 for displaying a user interface.
  • a suitable interface for monitoring the clustering analysis of Figure 4 includes a reference to the amount of data as a percentage of the entire database 10 that has been used in defining the model shown in Figure 6D. This interface allows the user to activate a button to suspend the operation of updating the model as well as adjusting the stopping criteria (Appendix D). The ability to suspend allows the database to be updated and then clustering can be resumed without resorting to a completely new analysis. This ability is particularly advantageous when clustering large databases where obtaining even a part of the data can take significant time.
  • FIGs 10 - 14 illustrate user interface screens that are depicted on a monitor 47 as data is clustered. These screens are illustrated in the example of the clustering framework described in this invention applied to scaling the K-means algorithm in particular. Scaling other clustering algorithms involves displaying potentially other relevant information concerning the model being constructed. Of course, this affects only display of quantities pertaining to specific model. General notions such as progress bar (302), information like 304, and buffer utilization 334, and 332 are independent of clustering algorithm and do not change with change in clustering algorithm.
  • this screen 300 illustrates a clustering process as that clustering takes place.
  • a progress bar 302 indicates what portion of the entire database has been clustered and a text box 304 above the progress bar 302 indicates how many records have been evaluated. In a center portion of the screen 300 two
  • graphs 3 10, 3 12 illustrate clustering parameters as a function of iteration number and cluster ID respectively.
  • the first graph 3 10 illustrates progress of the clustering in terms of iteration number which is displayed in the text box 314.
  • the iteration number refers to the number of data gathering steps that have occurred since clustering was begun.
  • an energy value for the clustering is calculated as defined in Appendix D method 2. As the clustering continues the energy decreases until a stopping criteria has been satisfied.
  • sixteen iterations are depicted.
  • the second graph 312 at the bottom of the screen is a graph of clustering parameters as a function of cluster number.
  • Covariance is defined from the model data ( Figure 6D) for a given cluster and a given dimension by the relation: SumSq/M - Sum*Sum/M 2
  • a plot of minimum covariance is therefore a plot of the dimension (l ...n) for a given cluster model having the least or minimum covariance.
  • a drop down list box 320 allows the user to select other indications of covariance.
  • a depiction of the dimension of the model having the maximum covariance ( Figure 10) for each of the ten clusters is shown in the bar graph 312.
  • An average covariance bar graph ( Figure 12) indicates the average of the covariance for each cluster over all cluster dimensions.
  • a different user selection via the drop down list box 320 ( Figure 13) shows the weight M for the ten clusters.
  • a dropdown list box 322 allows different cluster parameters such as model difference (Appendix D, method 1 ) to be plotted on the graph 3 10.
  • a row 326 of command buttons at the bottom of the screen allow a user to control a clustering process.
  • a parameter screen button 330 allows the user to view a variety of clustering parameters on a parameter screen (not shown). By accessing this screen the user can determine for example a maximum number of records or tuples that can be brought into memory to the data set RS in a given iteration. As an example, the user could indicate that this maximum value is 10,000 records.
  • data is summarized in DS, CS, and stored in RS. If a number of 10,000 records is chosen as the maximum,
  • the system limits the number of new data that can be read based upon the number of subclusters in the data set CS. Designate the number as ROWSMAX, then the amount of data records that can be currently stored in RS (Rscurrent) is ROWSMAX - 2*c where c is the number of subclusters in CS.
  • a progress bar 332 indicates a proportion of data buffers that are currently being used to store RS and CS datasets. This is also depicted as a percentage in a text box 334.
  • parameter screen also allows the user to define where to store the model upon completion of the clustering. If a process is suspended and the model is stored, the user can also use this screen to browse the computer disk storage for different previously stored models.
  • Text boxes 342, 344, 346, 348 in this panel indicate a number of subclusters c, and average, minimum and maximum variances for the subclusters using the above definition of variance.
  • a last text box 350 indicates the average size of the subclusters in terms of data points or tuples in the subclusters.
  • a stop button 360 stops the clustering and stores the results to disk.
  • a continue button 362 allows the process to be suspended and resumed by activating a resume button 364.
  • a generate batch button 366 allows the user to generate a clustering batch file which can be executed as a separate process.
  • Finally a close button 368 closes this window without stopping the clustering.
  • K- eans as well as many other members of the iterative clustering algorithms, are well-known to be extremely sensitive to initial starting condition. However, standard practice usually calls for trying multiple solutions from multiple random starting points.
  • One process for choosing multiple starting points is disclosed in copending patent application entitled “A method for refining the initial conditions for clustering with applications to small and large database clustering.” Which was filed in the United States Patent and Trademark Office on March 4, 1998 and is assigned to the assignee of the present invention.
  • the invention includes the ability to explore multiple models.
  • the key insights for this generalization are:
  • Retained points RS and the sets CS are shared amongst the all models;
  • Each model say Mi, will have its own discarded data sets DSM, (K sets, one for each cluster for each model) — if there are in models, there are a wxK discard sets; 2)
  • DSM discarded data sets
  • DSy The sufficient statistics for discarded data sets DSy for one of the models Mi are simply viewed as members of the global CS by all models other than Mi.
  • model Ml will "feel" the effect of this point twice: once in its own DSJ and another time when it updates over DS2 which will be treated as part of CS as far as MI is concerned. Similarly for M2. By entering in exactly one discard set, say DSJ, the point x still affects M2 when M2 updates over CS and counts DSJ as part of CS.
  • each pointer points to a different model data structure.
  • the model data structure is depicted in Figure 6D.
  • the structure CS and RS are shared by the multiple models.
  • An extended CLUSTERING procedure for the multiple model process takes into account the multiple model aspects of the structures DS and CS is perfromed on each of the S models.
  • DS structures for all the S models are determined by compressing data points into the appropriate DS data structure.
  • the CS data structures for each of the models are then determined from the points remaining in RS.
  • the CS sufficient statistics must be augmented with the sufficient statistics contained in the DS data structures of the other models.
  • a data point is compressed into DS, it enters the DS set of only one model at the step 240, hence there is no double counting of data
  • the multiple model analysis can be performed until one of the models satisfies the stopping criteria at the step 140.
  • An alternate system would continue to compute all the models until each model reaches a stopping criteria.
  • the scalable K- means process could be performed until a certain percentage of the models have reached a stopping criteria.
  • the multiple model implementation shares data structures between models and performs calculations on certain data unique to a given model. This analysis is susceptible to parallel processing on a computer 20 having multiple processing units 21
  • each of the data points that is accessed and used in the the K-means clustering procedure has the same weight as other data.
  • the process can allow suspension of a model build and resumption at a later time.
  • the stopping criteria can be reached and the model stored and later used for whatever purpose the user desires.
  • the database can be updated between clustering. In those instances it may be desirable to weight more recently added data more heavily than the data records used to cluster a database from a prior time. In this instance the recent records can be weighted by a heavier factor such as 1.5 per data point.
  • the weighting factor is applied to the data before updating the summaries.
  • the invention provides a way to add an additional attribute or dimension to the data, i.e. cluster number.
  • This additional attribute can be used as an index field for the data in the database. It can also be used as a pivoting attribute t ⁇ reorganize and present the data in a different fashion to the user of the database.
  • an exemplary data processing system for practicing the disclosed data mining engine invention includes a general purpose computing device in the form of a conventional computer 20, including one or more processing units 21, a system memory 22, and a system bus 23 that couples various system components including the system memory to the processing unit 21.
  • the system bus 23 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures.
  • the system memory includes read only memory (ROM) 24 and random access memory (RAM) 25.
  • ROM read only memory
  • RAM random access memory
  • the computer 20 further includes a hard disk drive 27 for reading from and writing to a hard disk, not shown, a magnetic disk drive 28 for reading from or writing to a removable magnetic disk 29, and an optical disk drive 30 for reading from or writing to a removable optical disk 3 1 such as a CD ROM or other optical media.
  • a hard disk drive 27 for reading from and writing to a hard disk, not shown
  • a magnetic disk drive 28 for reading from or writing to a removable magnetic disk 29
  • an optical disk drive 30 for reading from or writing to a removable optical disk 3 1 such as a CD ROM or other optical media.
  • hard disk drive 27, magnetic disk drive 28, and optical disk drive 30 are connected to the system bus 23 by a hard disk drive interface 32, a magnetic disk drive interface 33, and an optical drive interface 34, respectively.
  • the drives and their associated computer-readable media provide nonvolatile storage of computer readable instructions, data structures, program modules and other data for the computer 20.
  • the exemplary environment described herein employs a hard disk, a removable magnetic disk 29 and a removable optical disk 31 , it should be appreciated by those skilled in the art that other types of computer readable media which can store data that is accessible by a computer, such as magnetic cassettes, flash memory cards, digital video disks, Bernoulli cartridges, random access memories (RAMs), read only memories (ROM), and the like, may also be used in the exemplary operating environment.
  • a number of program modules may be stored on the hard disk, magnetic disk 29, optical disk 3 1, ROM 24 or RAM 25, including an operating system 35, one or more application programs 36, other program modules 37, and program data 38.
  • a user may enter commands and information into the computer 20 through input devices such as a keyboard 40 and pointing device 42.
  • Other input devices may include a microphone, joystick, game pad, satellite dish, scanner, or the like.
  • These and other input devices are often connected to the processing unit 21 through a serial port interface 46 that is coupled to the system bus, but may be connected by other interfaces, such as a parallel port, game port or a universal serial bus (USB).
  • a monitor 47 or other type of display device is also connected to the system bus 23 via an interface, such as a video adapter 48.
  • personal computers typically include other peripheral output devices (not shown), such as speakers and printers.
  • the computer 20 may operate in a networked environment using logical connections to one or more remote computers, such as a remote computer 49.
  • the remote computer 49 may be another personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the computer 20, although only a memory storage device 50 has been illustrated in Figure 1.
  • the logical connections depicted in Figure 1 include a local area network (LAN) 51 and a wide area network (WAN) 52.
  • LAN local area network
  • WAN wide area network
  • the computer 20 When used in a LAN networking environment, the computer 20 is connected to the local network 51 through a network interface or adapter 53. When used in a WAN networking environment, the computer 20 typically includes a modem 54 or other means for establishing communications over the wide area network 52, such as the Internet.
  • the modem 54 which may be internal or external, is connected to the system bus 23 via the serial port interface 46.
  • program modules depicted relative to the computer 20, or portions thereof may be stored in the remote memory storage device. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.
  • [tValue] ComputeT( alpha , DegreesFreedom): computes the t-statistics for the given value of alpha (1 -ConfidenceLevel, in our case) and DegreesFreedom (# of points in a cluster -1 in our case).
  • DSCopy( DSNew, DSOrig) copies DSOrig into DSNew, DSNew is altered, DSOrig remains the same.
  • [Weight Vec] ComputeWeightVec( DataPoint, Model ): for K-Means, returns the ⁇ 0, 1 ⁇ weight vector indicating the cluster to which DataPoint is assigned.
  • CVDiag Model(l).GetCVDiag()
  • TValue ComputeT( (alpha/n), Model(l).GetNum()-l ); // correct t- value
  • CVDiagonal 5 [integer] Floor(float) returns the integer obtained by rounding the float down to the nearest integer 6 DSCopy(DSNew, DSOrig) copies DSOrig into DSNew, DSNew is altered,
  • Models(I), Forj 1 , . ,r
  • [Model] VanillaKMean( Model , Data , StopTol ): takes initial values for Model and updates the model with the Data until the model ceases to changes, within StopTol. The updated model is returned in Model.
  • [CSMerged] MergeCS( CSEle l, CSElem2 ): take the sufficient statistics for 2 sub-clusters, CS l and CS2, and computes the sufficient statistics for the subcluster formed by merging CS l and CS2.
  • [SubModel] RandomSample( RSNew, kPrime ): randomly chooses kPrime elements from RSNew and uses these as the initial points for the vanilla k-mean algorithm. The elements are stored in SubModel.
  • [Weight Vec] ComputeWeightVec( DataPoint, Model ): computes the ⁇ 0,1 ⁇ weight vector with k elements. DataPoint is assigned to cluster j if the j-th element of the Weight Vec is 1 , the other elements are 0. 6. Append( integer, integerList ): appends the integer to the end of integerList.
  • RemoveCSCandidates( IndexList, CS ) removes the elements specified in IndexList from CS and returns the altered, smaller list in CS.
  • [BigCSList] AppendCS( CSList l, CSList2 ): creates the BigCSList by appending CSList2 to the end of CSList l .
  • [SubcPartner,SubcPartnerInd] FindMergePartner( Index, CS ): finds the element (subcluster) in CS that has center nearest to CS(Index) (a different subcluster) and returns this element in SubcPartner and its index.
  • [CSMergedElem] MergeCS(DataPoint,CSElem): merges a singleton data point into a CS cluster and returns the merged Cluster.
  • PointPerSubClust is an integer which used to determine the number of
  • MinPoints is an integer specifying the minimum number of points allowed in a
  • StopTol is a tolerance for the vanilla k-mean algorithm
  • vanilla k-Mean algorithm // choose the starting point for the vanilla k-Mean algorithm as kPrime random
  • CVDiag CSCandidates(l).GetCVDiag()
  • [CandMerge] MergeCS( CSMergePartner , CSNew( CSIndex ) );
  • PrevModelDiff is a vector of model deltas over the last (current-
  • OldMean OldModel(l).GetMean()
  • NewMean NewModel(l).GetMean()
  • NewDiff + Distance2Norm( OldMean , NewMean ); ⁇
  • NewDiff (NewDiff/k);
  • PrevEnergyDiff is a vector of energy deltas over the last (current- 1 ),..., (current-z) // iterations // StopTol is a scalar stopping tolerance
  • NewDiff abs(NewEnergy-OldEnergy),
  • UpdateModel (Model,RSNode,WeightVec) updates Model with singleton data point RSNode using weighting information specified in Weight Vec.
  • [dataSource] Open_DataSource(dataSourceName,scanType): opens up a connection to the given data source, from which all data points will be consumed.
  • [RS,numRowsRead] DataSource_ j getNextRows(dataSource,maxRSRows): retrieves up to maxRSRows from dataSource, using the appropriate scan type.
  • ModelsNew ModelsUpdate(Models,DSs,CS,RS,stopTol): see above description of ModelsUpdate.
  • [dataSource] Open_DataSource(dataSourceName); do ⁇
  • ModelsNew ModelsUpdate(Models,DSs,CS,RS,stopTol);

Abstract

In a data mining system (12), clusters are used to categorize data within each model. An initial set of estimates of the parameters of each model and each cluster are provided. A portion of the data in the database (10) is read from a storage medium and brought into a rapid access memory buffer (22). Data contained in the data buffer (22) is used to update the original guesses at the parameters of the model in each cluster over all models. Some of the data belonging to a cluster is summarized or compressed and stored as a reduced form of the data representing sufficient statistics of the data. If further data is needed to categorize the cluster, more data is gathered from the database (10) and used in combination with compressed data until a stopping criteria (140) is met.

Description

A Scalable System for Clustering of Large Databases
Field of the Invention
The present invention concerns database analysis and more particularly concerns an apparatus and method for clustering of data into groups that capture important regularities and characteristics of the data.
Background Art
Large data sets are now commonly used in most business organizations. In fact, so much data has been gathered that asking even a simple question about the data has become a challenge. The modern information revolution is creating huge data stores which, instead of offering increased productivity and new opportunities, are threatening to drown the users in a flood of information. Tapping into large databases for even simple browsing can result in an explosion of irrelevant and unimportant facts. Even people who do not 'own' large databases face the overload problem when accessing databases on the Internet. A large challenge now facing the database community is how to sift through these databases to find useful information.
Existing database management systems (DBMS) perform the steps of reliably storing data and retrieving the data using a data access language, typically SQL. One major use of database technology is to help individuals and organizations make decisions and generate reports based on the data contained in the database.
An important class of problems in the areas of decision support and reporting are clustering (segmentation) problems where one is interested in finding groupings (clusters) in the data. Clustering has been studied for decades in statistics, pattern recognition, machine learning, and many other fields of science and engineering. However, implementations and applications have historically been limited to small data sets with a small number of dimensions.
Each cluster includes records that are more similar to members of the same cluster than they are similar to rest of the data. For example, in a marketing application, a company may want to decide who to target for an ad campaign based on historical data about a set of customers and how they responded to previous campaigns. Other examples of such problems include: fraud detection, credit approval, diagnosis of system problems, diagnosis of manufacturing problems, recognition of event signatures, etc. Employing analysts (statisticians) to build cluster models is expensive, and often not effective for large problems (large data sets with large numbers of fields). Even trained scientists can fail in the quest for reliable clusters when the problem is high-dimensional (i.e. the data has many fields, say more than 20).
Clustering is a necessary step in the mining of large databases as it represents a means for finding segments of the data that need to be modeled separately. This is an especially important consideration for large databases where a global model of the entire data typically makes no sense as data represents multiple populations that need to be modeled separately. Random sampling cannot help in deciding what the clusters are. Finally, clustering is an essential step if one needs to perform density estimation over the database (i.e. model the probability distribution governing the data source). Applications of clustering are numerous and include the following broad areas: data mining, data analysis in general, data visualization, sampling, indexing, prediction, and compression. Specific applications in data mining including marketing, fraud detection (in credit cards, banking, and telecommunications), customer retention and churn minimization (in all sorts of services including airlines, telecommunication services, internet services, and web information services in general), direct marketing on the web and live marketing in Electronic Commerce. Clustering is an important area of application for a variety of fields including data mining, statistical data analysis, compression, and vector quantization. Clustering has been formulated in various ways. The fundamental clustering problem is that of grouping together (clustering) data items that are similar to each other. The most general approach to clustering is to view it as a density estimation problem. We assume that in addition to the observed variables for each data item, there is a hidden, unobserved variable indicating the "cluster membership" of the given data item. Hence the data is assumed to arrive from a mixture model and the mixing labels (cluster identifiers) are hidden. In general, a mixture model M having K clusters Ci, i=l,...,K,
assigns a probability to a data point x as follows: Pr(x | M) = Wt - Pr(x | Ci,M) ι=l where W, are called the mixture weights. The problem then is estimating the parameters of the individual Ci. Usually it is assumed that the number of clusters K is known and the problem is to find the best parameterization of each cluster model. A popular technique for estimating the model parameters (including cluster parameters and mixture weights) is the EM algorithm (see P. Cheeseman and J. Stutz, "Bayesian Classification (AutoClass): Theory and Results", in in Advances in Knowledge Discovery and Data Mining, Fayyad, U., G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy( Eds.), pp. 153-180. MIT Press, 1996; and A.P. Dempster, N.M. Laird, and D.B. Rubin, "Maximum Likelihood from Incomplete Data via theEM algorithm". Journal of the Royal statistical Society, Series B, 39(1): 1-38, 1977). There are various approaches to solving the optimization problem of determining (locally) optimal values of the parameters given the data. The iterative refinement approaches are the most effective. The basic algorithm goes as follows:
1. Initialize the model parameters, producing a current model. 2. Decide memberships of the data items to clusters, assuming that the current model is correct.
3. Re-estimate the parameters of the current model assuming that the data memberships obtained in 2 are correct, producing new model.
4. If current model and new model are sufficiently close to each other, terminate, else go to 2.
The most popular clustering algorithms in the pattern recognition and statistics literature belong to the above iterative refinement family: the K-Means algorithm J. MacQueen, "Some methods for classification and analysis of multivariate observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability. Volume I, Statistics, L. M. Le Cam and J. Neyman (Eds.). University of California Press, 1967. There are many variants of these that iteratively refine a model by rescanning the data many times. These algorithms have found many applications recently, including in industry and science. The difference between the EM and K-Means is the membership decision (step 2). In K-Means, a data item belongs to a single cluster, while in EM each data item is assumed to belong to every cluster but with a different probability. This of course affects the update step (3) of the algorithm. In K-Means each cluster is updated based strictly on its membership. In EM each cluster is updated by the entire data set according to the relative probability of membership.
Summary of the Invention The invention represents a methodology for scaling clustering algorithms to large databases. The invention enables effective and accurate clustering in one or less scans of a database. Use of the invention results in significantly better performance than prior art schemes that are based on random sampling. These results are achieved with significantly less memory requirement and acceptable accuracy in terms of approaching the true solution than if one had run the clustering algorithm on the entire database.
Known methods can only address small databases (ones that fit in memory) or resort to sampling only a fraction of the data. The disclosed invention is based on the concept of retaining in memory only the data points that need to be present in memory. The majority of the data points are summarized into a condensed representation that represents their sufficient statistics. By analyzing a mixture of sufficient statistics and actual data points, significantly better clustering results than random sampling methods are achieved and with similar lower memory requirements. The invention can typically terminate well before scanning all the data in the database, hence gaining a major advantage over other scalable clustering methods that require at a minimum a full data scan.
The invention concerns a framework that supports a wide class of clustering algorithms. The K-means algorithm as an example clustering algorithm that represents one specific embodiment of this framework. The framework is intended to support a variety of algorithms that can be characterized by iteratively scanning data and updating models. We use K-Means since it is well-known and established clustering method originally known as Forgy's method and has been used extensively in pattern recognition. It is a standard technique for clustering, used in a wide array of applications and even as way to initialize the more expensive EM clustering algorithm. When working over a large data store, one needs to pay particular attention to certain issues of data access. A clustering session may take days or weeks, and it is often desirable to update existing models as data arrives. A list of desirable data mining characteristics follows: The invention satisfies all these:
1. Clustering should run within one scan (or less) of the database if possible: a single data scan is considered costly, early termination if appropriate is highly desirable.
2. On-line "anytime" behavior: a "best" answer is always available from the system, with status information on progress, expected remaining time, etc.
3. Suspendable, stoppable, resumable; incremental progress saved to resuming a stopped job. 4. An ability to incrementally incorporate additional data with existing models efficiently.
5. Should work within confines of a given limited RAM buffer
6. Utilize variety of possible scan modes: sequential , index, and sampling scans if available. 7. Should have the ability to operate with forward-only cursor over a view of the database. This is necessary since the database view may be a result of an expensive join query, over a potentially distributed data warehouse, with much processing required to construct each row (case).
The technique embodied in the invention relies on the observation that clustering techniques do not need to rescan all the data items as it is originally defined and as implemented in popular literature and statistical libraries and analysis packages. The disclosed process may be viewed as an intelligent sampling scheme that employs some theoretically justified criteria for deciding which data can be summarized and represented by a significantly compressed set of sufficient statistics, and which data items must be carried in computer memory, and hence occupying a valuable resource. On any given iteration of the invention, we partition the existing data samples intro three subsets: A discard set (DS), a compression set (CS), and a retained set (RS). For the first two sets, we discard the data but keep representative sufficient statistics that summarize the subsets. The last, RS, set is kept in memory. The DS is summarized in a single set of sufficient statistics. The compression set CS is summarized by multiple sufficient statistics representing subclusters of the CS data set. The invention operates by obaining a next available (possibly random) sample from a database to fill free space in buffer. A current model of the clustering is then updated over the contents of the buffer. Elements of the new sample are identified to determine whether they need to be retained in the buffer (retained set RS); they can be discarded with updates to the sufficient statistics (discard set DS); or they can be reduced via compression and summarized as sufficient statistics (compression set CS). Once this has been done a determination is made to see if a stopping criteria is satisfied. If so terminate clustering, if not then sample more data.
The exemplary embodiment satisfies the above-mentioned important issues faced during data mining. A clustering session on a large database can take days or even weeks. It is often desirable to update the clustering models as the data arrives and is stored. It is important in this data mining environment to be able to cluster in one scan (or less) of the database. A single scan is considered costly and clustering termination before one complete scan is highly desirable. An exemplary embodiment of the invention includes a model optimizer. A multiple number of different clustering models are simultaneously generated in one or less scans of the database. The clustering analysis stops when one of the models reaches a stopping criteria. Alternately, the clustering can continue until all of the multiple models are complete as judged by the stopping criteria. These and other objects, advantages and features of the invention will be better understood from a detailed description of an exemplary embodiment of the invention which is described in conjunction with the accompanying drawings.
Brief Description of the Drawings Figure 1 is a schematic illustration of a computer system for use in practicing the present invention;
Figure 2 is schematic depiction of a clustering process of the invention wherein multiple clustering models are obtained during a single scan of the a database;
Figure 3 A and 3B are schematic depictions showing software components of the invention;
Figure 4 is a flow diagram of an exemplary embodiment of the invention; Figure 5 is a one-dimensional plot of data distribution of three clusters of data; Figure 6 A - 6D are illustrations of data structures utilized for storing data in accordance with the exemplary embodiment of the invention;
Figure 7 is a flow diagram of an exemplary embodiment of an extended K- means analysis of data;
Figure 8 is a data structure for use in determining multiple data models through practice of the exemplary embodiment of the present invention;
Figure 9 is a plot of data points in two dimensions showing a clustering of data; Figures 10 - 14 are user interface screens indicating various aspects of the clustering process.
Detailed Description of Exemplary Embodiment of the Invention
A large database 10 for which the present invention has particular utility contains many records stored on multiple, possibly distributed storage devices. Each record has many attributes or fields which for a representative database might include age, income, number of children, number of cars owned etc. A goal of the invention is to characterize clusters of data in the database 10. This task is straightforward for small databases (all data fits in the memory of a computer for example) having records that have a small number of fields or attributes. The task becomes very difficult, however, for large databases having huge numbers of records with a high dimension of attributes.
Overview of Scalable Clustering Figure 4 is a flow chart of the process steps performed during a scalable clustering analysis of data in accordance with the present invention. An initialization step 100 includes a step of initializing a number of data structures shown in Figure 6A - 6D and choosing a cluster number K for characterizing the data.
A next step 1 10 is to sample a portion of the data in the database 10 from a storage medium to bring that portion of data within a random access memory (into RAM for example, although other forms of random access memory are contemplated) of the computer 20 shown in Figure 1. In general, the data has a large number of fields so that instead of a single dimension analysis, the invention characterizes a large number of vectors where the dimension of the vector is the number of attributes of the data records in the database. The data structure 180 for this data is shown in Figure 6C to include a number r of records having a potentially large number of attributes. The gathering of data can be performed in either a sequential scan that uses only a forward pointer to sequentially traverse the data or an indexed scan that provides a random sampling of data from the database. When using the index scan it is a requirement that data not be accessed multiple times. This can be accomplished by marking data tuples to avoid duplicates, or a random index generator that does not repeat. In particular, it is most preferable that the first iteration of sampling data be done randomly. If it is known the data is random within the database then sequential scanning is acceptable. If it is not known that the data is randomly distributed, then random sampling is needed to avoid an inaccurate representative of the database.
Returning to Figure 4, a processor unit 21 of the computer 20 next executes 120 a clustering procedure using the data brought into memory in the step 110 as well as compressed data in the CS and DS data structures. In accordance with an exemplary clustering process described in greater detail below, the processor 21 assigns data contained within the portion of data brought into memory to a cluster and determines a set of parameters for each attribute of the data assigned to a given cluster. A data structure for the results or output model of the analysis is a model of the clustering of data and is depicted in Figure 6D. This model includes K records, one for each cluster. In an exemplary embodiment (for supporting the K-means clustering algorithm) which assumes that the attributes of the database are independent of each other, each record has three required components: 1 ) a vector 'Sum' representing the sum for each of the attributes or dimensions of the vector data records in a given cluster, 2) a vector 'Sumsq' representing the sum of the attributes squared and 3) an integer 'M' counting the number of data records contained in or belonging to the corresponding cluster. These are sufficient to compute the mean (center) and covariance (spread) of the data in a cluster. Note that these structures are maintained in the case of K-means because these are the set of model parameters that characterize K-means clustering. In general, these data structures need to hold the parameters describing the best model. In case of discrete data, these structures would hold parameters of the multinomial distribution, or some other distribution on discrete data. In case of EM with mixture of Gaussians, these structures would hold the means and covariance matrices of the clusters. Note that although Figure 6 (A-D) shows SUM and SUMSQ structures, some of these structures change depending on the clustering algorithm being used. In case of EM, the SUMSQ vector (array) may be replaces with a matrix holding covariances of data in the cluster. The parameters represented in the structures enable the model to assign a probability of membership for every data item in every cluster. Our scalable framework needs this probability to determine data item membership in DS, CS, and RS sets (discussed later), as part of its data compression stage.
A next step 130 in the Figure 4 flowchart summarizes at least some of the data used in the present iteration to characterize the K clusters. This summarization is contained in the data structures DS, CS of Figures 6A and 6B. The summarization takes up significantly less storage in a computer memory 25 than the vector data structure (Figure 6D) needed to store individual records. Storing a summarization of the data in the data structures of Figures 6 A and 6B frees up more memory allowing additional data to be sampled from the database 10. Before looping back to get more data the processor 21 determines 140 whether a stopping criteria has been reached. One stopping criteria that is used is whether the analysis has produced a good enough model (Figure 6D) by a standard that is described below. A second stopping criteria has been reached if all the data in the database 10 has been used in the analysis. One important aspect of the invention is the fact that instead of stopping the analysis, the analysis can be suspended. Data in the data structures of Figure 6A - 6D can be saved (either in memory or to disk) and the scalable clustering analysis can then be resumed later. This allows the database to be updated and the analysis resumed to update the clustering statistics without starting from the beginning. It also allows another process to take control of the processor 21 without losing the state of the clustering analysis. The suspension could also be initiated in response to a user request that the analysis be suspended by means of a user actuated control on an interface presented to the user on a monitor 47 while the Clustering analysis is being performed. Figures 3 A and 3B depict an operating environment of the invention. Tjis is shown also in NEWFIG. Data in the database 10 is accessed through a database management system that sends data to a data mining engine 12. The data mining engine 12 processes requests from an application 14 and responds by sending back model summaries to the application. The software components of the data mining engine 12 are depicted in more detail in Figure 3 A.
K-Means Clustering There is a large literature of known data clustering techniques. One known technique is the so called K-means clustering process which is summarized in Duda- Hart (Pattern Classification and Scene Analysis) 1973 John Wiley & Sons, Inc., New York. An exemplary embodiment of the present scalable clustering analysis is described by reference to the K-Means Clustering process.
Consider a two dimensional depiction of data from a database illustrated in Figure 9. Spaced across the two dimensional space of the Figure are a number of data points. In a typical database this would be a depiction of two attributes of a multi- attribute record..
One can visually determine that the data in Figure 9 is lumped or clustered together. Classifying the data into clusters is dependent on a staring cluster number. If one chooses three clusters for the data of Figure 9 the data would generally fall into the clusters K l, K2, K3 of the figure.
The K-Means algorithm takes as input: the number of clusters K, a set of K initial estimates of the cluster means, and the data set to be clustered. The means
(centroids) define the parameters of the model. One traditional K-means evaluation starts with a random choice of cluster centroids or means that are randomly placed within the extent of the data on the x axis. Call these M l, M2, and M3 in Figure 9.
Each cluster is represented by its mean and each data item is assigned membership in the cluster having the nearest mean. Distance is the Euclidean distance (or L2 norm): for a data point (d-dimensional vector) x and mean μ, is given by:
Dist(x,//) = ^x", ~μ, Ϋ ■ The cluster model is updated by computing the mean over
the data items assigned to it. The model assumptions relating to the classic K-Means algorithm are: 1 ) each cluster can be effectively modeled by a Gaussian distribution with diagonal covariance matrix having equal diagonal elements (over all clusters), 2) the mixture weights (WJ are also assumed equal. Note that K-Means is only defined over numeric (continuous-valued) data since the ability to compute the mean is a requirement. A discrete version of K-Means exists and is sometimes referred to as
K) harsh EM . The K-Means algorithm finds a locally optimal solution to the problem of minimizing the sum of the L2 distance between each data point and its nearest cluster center (usually termed "distortion") .
For a database that fits in memory all data within the database can be used to calculate the K-means clustering centroids. The output from such a prior art process will be the K centroids and the number of data points that fall within each of the K clusters.
In accordance with the present invention, the data from the database 10 is brought into a memory 22 (Figure 1) and once the K-means calculation is performed it is assigned to one of three data sets. A retained data set (RS) is kept in memory 22 for further use in performing the K-means analysis. A so called discard data set (DS) and a compressed data set (CS) are summarized in the form of sufficient statistics that are retained in memory. For a random sampling of data the more data that is gathered the better the clustering calculation becomes. For a process in accordance with the present invention the estimates do get better as more data is sampled, but the estimates are accurate even though memory requirements limit sampling to a low percentage of the entire data base.
Data Structures Figures 6A - 6D summarize the data structures used to perform the scalable K- means analysis. The output from the analysis is stored in a data structure designated MODEL (Figure 6D) which includes an array 152 of pointers, each pointer points to a first vector 154 of n elements (floats) 'Sum', a second vector 156 of n elements (floats) 'SumSq', and a scalar 158 designated M. The number n corresponds to the number of attributes of the database records that are being clustered.
Figure 6A depicts a data structure designated DS including an array 160 of pointers, each of which identifies a vector 162 of n elements (floats) 'Sum' , a vector 164 of n elements (floats) 'SumSq', and a scalar 166 designated 'M'.
A further data structure designated CS is an array of c pointers 170, where each pointer points to a vector 172 of n elements (floats) 'Sum', a vector 174 of n elements (floats) 'SumSq', and a scalar 176 designated as M.
1 1 An additional data structure designated RS (Figure 6C) is an array 180 of r elements where each element points a vector of n elements (floats) representing a singleton data point of a type designated SDATA. As data is read in from the database it is stored in the set RS and this data is hot associated with any of the K clusters. An exemplary implementation of the scalable K-means analysis has RS being an array of pointers to elements of type SDATA, and an associated SumSq vector is null and the scalar M = 1.
In an exemplary embodiment of the invention, multiple models are simultaneously generated. This structure is depicted in Figure 8 wherein multiple models are pointers to arrays of models such as the data structures shown in Figure 6D.
Table 1 below is a list often SDATA vectors which constitute sample data from a database 10 and are stored as individual vectors in the data structure RS.
TABLE
CaselD AGE INCOME CHILDREN CARS
1 30 40 2 2
2 26 21 0 1
3 18 16 0 1
4 45 71 2
5 41 73 2 3
6 67 82 6 3
7 75 62 4 1
8 21 23 1 1
9 45 5 1 3 2
Figure imgf000014_0001
10 28 19 0 0
Table 2 below tabulates mean values chosen for a starting point assuming K=3 for performing the scalable K-means clustering analysis on the data of table 1.
TAI 3LE 2
Cluster # AGE INCOME CHILDREN CARS
1 55 50 2.5 2
2 30 38 1.5 2
Figure imgf000014_0002
3 20 24 1 1
12 An important concept of the present invention is the summarization or compression of data points of type SDATA contained in the dataset RS (TABLE 1) sampled from the database into the two data structures DS, CS to allow more sampling of data from the database. During each processing iteration of the Figure 4 flowchart the scalable K-means analysis calls an 'extended K-means' procedure 120 that utilizes the compressed data as well as the remaining data samples contained in the dataset RS.
On a first iteration through the Figure 4 process the set DS (Figure 6A) is empty. Updates to the set DS are performed at the step 130 for each cluster of the range, i = 1,..., K. For each cluster i the processor determines which singleton data elements (elements of the set RS of type SDATA), assigned to cluster i will not change cluster membership over future data samples. These points will be used to augment the i-th element of the set DS which contains the sufficient statistics summarizing these singleton points. These points are removed from the set RS and used to update the sufficient statistics for the i-th cluster of the set DS.
Two conceptual data structures help explain a first embodiment of the method of updating the set DS. This first embodiment is referred to as the Worst Case Analysis method. These conceptual data structures hold the upper and lower bounds defining an n-dimensional (n = # attributes) confidence interval (Cl) on the parameters of the model (in case of K-means the parameters are the means or centroids of the K clusters) computed so far. A list structure designated LOWER is a vector of k elements (one for each cluster) where each element points to a vector of n elements (floats) holding the lower bounds for each attribute of the CI on the mean of the corresponding cluster. For example LOWER(3).LowVec(2) is the value of the lower bound on the CI for the third cluster along dimension 2. A second structure designated UPPER is a vector of K elements (one for each cluster) where each element points to a vector of n elements (floats) holding the upper bounds for the CI on the parameters of the model (mean or centroid in case of K-means) of the corresponding cluster. Singleton Points (Elements of RS) not changing cluster assignment when the K cluster centers are perturbed, within their respective confidence intervals in a worst-case fashion, can be summarized by adding them to the set DS and removing them from RS. Appendix A is a summarization of the Worst Case Analysis that defines LOWER and UPPER as well as the process of updating of the set DS using the Worst Case Analysis.
A second embodiment of the process of updating the dataset DS is referred to as the Threshold Analysis. For this second embodiment a data structure is used that helps sort the singleton elements of RS (of type SDATA) by their Mahalanobis distance ( See Duda and Hart, Pattern Classification and Scene Analysis referenced above) to a given cluster center. A structure RSDist is a list of r elements (r is the number of singleton data elements in the set RS) where each element in RSDist points to 2 objects: 1 ) float called "MahalDist" which holds the Mahalanobis distance of the corresponding element in RS to the nearest cluster center and 2) an integer indicating the cluster whose center is nearest to the given point in RS, called "Oust Assign". Appendix C summarizes the calculation of the Mahalanobis distances.
NOTE that Mahanalobis Distance is used in case of K-means as an example. In general, we require the ability to measure the distance of a point to a cluster. This distance can be obtained from any model as the probability that a data item belongs to the cluster. Low probability data items are FAR, while high probability items are CLOSE according to this distance. Our framework for scaling clustering in general requires ability to extract this notion of "distance". This is always possible, regardless of model being used. The Mahalanobis distance in case when clusters are represented by multivariate Gaussians is directly proportional to probability assigned to a data item by the Gaussian in that cluster.
A third embodiment for updating the dataset DS is based upon the distance of a data point from the then current parameters characterizing the cluster (e.g. the cluster mean in K-means). If this distance for a point rs contained in RS is less than a specified distance, then the point is removed from RS and added to DS by updating the sufficient statistics of DS to include the rs point.
Figure 5 depicts a clustering in one dimensional, for example, income data from the table 1 list of data. Three clusters K l, K2, K3 are made up of much larger numbers of data vectors SDATA. The data structure DS that summarizes data for the cluster designated K 1 is centered within the generally Gaussian shaped curve depicting that cluster. Regardless of the technique used, certain data safely 'belongs' to the
14 cluster K 1 so it is safe to compress or summarize the data in this range in the form of the sufficient statistics contained in the DS (discard region) data structure.
Subclusters in dataset CS After the compression of data into the DS data structure, there is still a fairly large amount of data (SDATA) contained to the left and the right of the centroid of Kl in Figure 5 that neither the worst case analysis (Appendix A) nor the threshold analysis (Appendix B) identifies for compression into the set DS. These point fall within the 'compression region' for cluster 1. The present invention also summarizes at least some of this data in the form of a dataset designated CS. An exemplary process for determining the CS dataset is summarized in the psuedocode of Appendix C. Briefly, a dataset RSNew is made up of the dataset RS after removing the set DS from the original set RS.
The process of determining the new CS data set begins by merging and removing any singleton points from RSNew into CS which can safely be merged without violating a specified "density" criteria. For each data point in RSNew, the appendix C procedure finds the CS subcluster which is closest to it. If the data point can be merged into CS without violating the specified "density criteria" of the CS subcluster, then the data point is merged into that particular CS subcluster and removed from RSNew. If it doesn't satisfy the criteria, then the data point is left in RSNew.
The process of determining the data set CS then continues by finding a set of "dense" subclusters within the set RSNew. This is done by performing a traditional K- means analysis on the data in the set RSNew using a cluster number K' (Kprime) greater than K, the number of clusters used in performing the scalable K-means analysis. The set of sufficient statistics (Sum, Sumsq, and M) for the K' subclusters found by this procedure are appended to the current set of sufficient statistics in the dataset CS. Hence CS is augmented by K' elements to produce a number c of subclusters. This augmented list is then filtered and elements are merged (if possible), reducing the size of the list. Note that regardless of clustering algorithm being used in the main scalable framework (i.e. the clustering algorithm being scaled to large DB ), the clustering algorithm used to determine CS can be different. K-means is disclosed for secondary clustering (determining CS), but other clustering algorithms can be
15 employed [including hierarchical agglomerative clustering or any other clustering technique to find local structure in CS].
If the number of data points (M in the data structure CS) is less than a threshold value, (MinPoints in Appendix C) the data from this analysis is not clustered and the data is instead kept in the set RSNew. Furthermore only dense clusters are kept in CS. For each subcluster remaining after the threshold value of points criteria has been checked, if the maxiumum measure of spread computed from (SUMSQ) along any of the n dimensions (attributes) of the candidate subcluster is less than a threshold (StdTol in appendix C) the data from that subcluster is also left in RSNew and not summarized in CS. These two criteria remove the elements of CS corresponding to subclusters which are either too small (in terms of number of points) or too 'spread out'
A final step is the combining of subclusters using hierarchical agglomerative clustering. An attempt is made to merge two elements of the dataset CS. If the larger, merged subcluster still satisfies the 'spread' criteria discussed above, these two subclusters are removed from the set CS, and the larger subcluster representing these two smaller subclusters is added to the dataset CS. This process continues until the merging process produces no larger subclusters that still satisfy the "dense" criteria. The data structure CS contains c elements that are then used in the K-means analysis.
Extended CLUSTERING procedure An extended CLUSTERING procedure 120 includes looping constructs for updating the current model that are summarized in the flow chart of Figure 7. Certain utility functions are needed to perform this extended CLUSTERING procedure. A function ModelCopy( ModelNew , ModelOrig ) copies the model ModelOrig into
ModelNew. A function Length( DataStructure ) returns the length of the pointer array for the data structures of Figure 6 so that for example, Length(CS) = c and Length(RS) = r. Zero( Model ) takes the data structure for the model in Figure 6D and sets all elements to 0.0. A function Distance2Norm( point 1, point2) measures the distance between two vectors point 1 and point2.
The extended K-means procedure of Figure 7 begins with a step 200 of copying the existing model into an old model data structure. The process next
16 determines 202 the length of the RS, DS, and CS data structures of Figures 6A - 6C and returns the values k, c, and r. The data structure NewModel is then zeroed or initialized 204. The process updates the NewModel until a test 206 indicates a stopping criteria has been met. If the stopping criteria has not been met, the process saves 208 the new model in the old model data structure and branches 210 to zero a next subsequent new model. The test 206 is similar to the test 140 described in conjunction with the scalable K-means process overview.
After initialization of a new model, that model is updated in a loop 220 that updates the model for each of the r vectors in the dataset RS. The loop gathers data 222 a point at a time from the set RS and determines 224 what cluster to assign that data point to. This is done by finding the distance from the data point to the mean of each of the then existing means of the old model. By reference to Figure 6D (for particular example of sclaing K-means, hence EXTENDED K_MEANS algorithm) it is seen that the model includes the Sum for a given cluster K and therfore the mean or centroid of each dimension is given by this value divided by the scalar M for the cluster. Once the closest cluster is found, the New model Sum and SumSq component for that closest cluster is updated by adding the data point to the vector Cluster(closest).sum and then squaring the components and adding them to the Cluster(closest). SumSq vector. The scalar M for a cluster is incremented by one for each point added to that cluster.
Once the loop over the r vectors is completed, the procedure updates the model based on the compressed statistics stored in the c subclusters found in the data structure CS. On an initial loop through the Figure 4 scalable K-means process there are no CS or DS structures. Table 3 below indicates the contents of the model (Figure 6D) after the RS portion of the extended K-means process on data points of Table 1.
TABLE 3
SUM
Cluster # AGE INCOME CHILDREN CARS M
1 228 288 15 9 4
2 75 91 5 4 2
3 93 79 1 3 4
SUMSQ
Figure imgf000019_0001
Cluster # AGE INCOME CHILDREN CARS
17 1 5 1984 8244 225 81
2 5625 8281 25 16
Figure imgf000020_0001
3 8649 6241 1 9
The ten records of TABLE 1 will fit in memory, and a conventional K-means analysis is possible. For a large data base containing millions of records, the ten records constitute only a part of one data gathering step. Table 4 below lists a K- means clustering of data performed on the ten records with K = 3.
TABLE 4
Cluster # AGE INCOME CHILDREN CARS
1 57 72 3.75 2.25
2 37.5 45.5 2.5 2
Figure imgf000020_0002
3 23.25 19.75 0.25 0.75
To free up computer memory for gathering more of the millions of records some of the ten records shown in Table 1 are candidates to be compressed into the data structures CS, DS. The cluster averages for the income attribute of the ten records are labeled in the Figure 5 depiction. Record number 10 has an income of '19' and for this one dimension falls safely within the DS (discard region) centered around the cluster Kl in Figure 5. Visualizing the situation over many attributes becomes more difficult but the techniques summarized in the appendices deal with vectors and identify records within RS for compression. Record number 8 in Table 1 has an income of '23 '. Assume this record does not fall within the DS region and therefore becomes a candidate for inclusion in the CS (compress) dataset. Note, the cluster mean for the second cluster K2 is at an income of 45.5k dollars. Data falling between the two means of 19.75 and 45.5 typically will not be classed in either the DS or the CS dataset. It is retained in RS and used on the next iteration to perform the clustering.
After the initial iteration of the Figure 4 process, the CS and DS structures contain sufficient statistics, and the extended clustering procedure of Figure 7 must take into account this data in determining the new model when the procedure is called at the step 120 of the scalable K-means analysis. To update the model based on the sufficient statistics contained in the dataset CS the Figure 7 procedure executes a loop 230 over each of the subclusters c in CS and determines which of the K clusters in the Model (Figure 6D) that subcluster is closest to. Assume subcluster p is closest to cluster q. When this fact is discovered the sufficient statistics of cluster q are updated by adding the contents subcluster(p).sum to cluster(q).sum and the statistics subcluster(p).sumsq to cluster(q).sumsq. Additionally, the value of subcluster(p).M for the subcluster is added to the to the value cluster(q).M.
At a step 240 the extended CLSUTERING procedure updates the NewModel for the clusters summarized in DS. There is no need to search for the cluster nearest the clusters in DS since the elements of DS will always (or are assumed to be always) assigned to the same cluster. The step 240 merely loops over the clusters in DS and adds their sufficient statistics to the new model of figure 6D. In case of K-means example: (NewModel(l).Sum += DS(l).Sum, NewModel(l). SumSq += DS(l).SumSq and NewModel(l).M += DS(l).M) Once the contributions of CS and DS are added the stopped criteria is checked 206 to see if the procedure has converged to a solution. In one exemplary embodiment a variable CenterDist is set to zero and for each of the clusters K, a distance between the centroid of the old model and the centroid of the new model is determined and added to the CenterDist variable. Once all K distances have been calculated, and added together the CenterDist value is divided by the number of clusters K and compared to a value 'StopTol' which is used as a measure of how stable the model has become. If the value of CenterDist is smaller than the value 'StopTol' then the procedure returns, otherwise the procedure branches back to recalculate the model using the same data in RS, CS, and DS but with a different "old model".
Stopping criteria at the step 140
Each time the procedure 120 returns, the RS, DS and CS data structures are updated and the test of the stopping criteria 140 is performed. Three alternative stopping criteria are proposed for use in the scalable K-mean procedure, (methods 1 and 2 of the Appendix D pseudocode summarize two of these criteria) A first method terminates the analysis if the difference between the K- means, measured in some norm, over a given number of data samples is below a given tolerance. A second method terminates if the difference in an "energy" function (measure of distortion)
19 minimized by the k-mean analysis falls below a given tolerance over a fixed number of data samples. A third terminates if the number of data samples from the database is exhausted. A fourth stopping criteria is actually a suspension of the scalable K-means analysis rather than stopping. We note that if storage permits, the most general storage scheme would keep in main memory the last z models, hence easily allowing the plug-in of either stopping criteria 1 or 2 by easily computing either PrevModelDiff (in the case that the first stopping criteria is chosen , (see appendix D) from these z models or by computing PrevEnergyDiff (in the case that the second stopping criteria is chosen, see appendix D).
As seen in Figure 1 the computer 20 includes a monitor 47 for displaying a user interface. A suitable interface for monitoring the clustering analysis of Figure 4 includes a reference to the amount of data as a percentage of the entire database 10 that has been used in defining the model shown in Figure 6D. This interface allows the user to activate a button to suspend the operation of updating the model as well as adjusting the stopping criteria (Appendix D). The ability to suspend allows the database to be updated and then clustering can be resumed without resorting to a completely new analysis. This ability is particularly advantageous when clustering large databases where obtaining even a part of the data can take significant time. User Interface
Figures 10 - 14 illustrate user interface screens that are depicted on a monitor 47 as data is clustered. These screens are illustrated in the example of the clustering framework described in this invention applied to scaling the K-means algorithm in particular. Scaling other clustering algorithms involves displaying potentially other relevant information concerning the model being constructed. Of course, this affects only display of quantities pertaining to specific model. General notions such as progress bar (302), information like 304, and buffer utilization 334, and 332 are independent of clustering algorithm and do not change with change in clustering algorithm. Turning to Figure 9, this screen 300 illustrates a clustering process as that clustering takes place. A progress bar 302 indicates what portion of the entire database has been clustered and a text box 304 above the progress bar 302 indicates how many records have been evaluated. In a center portion of the screen 300 two
20 graphs 3 10, 3 12 illustrate clustering parameters as a function of iteration number and cluster ID respectively. The first graph 3 10 illustrates progress of the clustering in terms of iteration number which is displayed in the text box 314. The iteration number refers to the number of data gathering steps that have occurred since clustering was begun. In the Figure 9 depiction an energy value for the clustering is calculated as defined in Appendix D method 2. As the clustering continues the energy decreases until a stopping criteria has been satisfied. In the graph 310 of Figure 9 sixteen iterations are depicted.
The second graph 312 at the bottom of the screen is a graph of clustering parameters as a function of cluster number. In the depiction shown there are ten clusters (shown in the text box 3 16) and the minimum covariance for each of these ten clusters is shown. Covariance is defined from the model data (Figure 6D) for a given cluster and a given dimension by the relation: SumSq/M - Sum*Sum/M2 A plot of minimum covariance is therefore a plot of the dimension (l ...n) for a given cluster model having the least or minimum covariance. A drop down list box 320 allows the user to select other indications of covariance. By selecting a maximum for this parameter, a depiction of the dimension of the model having the maximum covariance (Figure 10) for each of the ten clusters is shown in the bar graph 312. An average covariance bar graph (Figure 12) indicates the average of the covariance for each cluster over all cluster dimensions. A different user selection via the drop down list box 320 (Figure 13) shows the weight M for the ten clusters. In a similar manner, a dropdown list box 322 allows different cluster parameters such as model difference (Appendix D, method 1 ) to be plotted on the graph 3 10. A row 326 of command buttons at the bottom of the screen allow a user to control a clustering process. A parameter screen button 330 allows the user to view a variety of clustering parameters on a parameter screen (not shown). By accessing this screen the user can determine for example a maximum number of records or tuples that can be brought into memory to the data set RS in a given iteration. As an example, the user could indicate that this maximum value is 10,000 records.
As outlined above, as the clustering process is performed data is summarized in DS, CS, and stored in RS. If a number of 10,000 records is chosen as the maximum,
21 the system limits the number of new data that can be read based upon the number of subclusters in the data set CS. Designate the number as ROWSMAX, then the amount of data records that can be currently stored in RS (Rscurrent) is ROWSMAX - 2*c where c is the number of subclusters in CS. A progress bar 332 indicates a proportion of data buffers that are currently being used to store RS and CS datasets. This is also depicted as a percentage in a text box 334.
Other parameters that can be modified on the parameter screen are choice of the stopping tolerance, choice of the stopping procedure, choice of parameters for combining subclusters and adding data points to subclusters, and choice of the compression procedure used to determine DS data set candidates. The parameter screen also allows the user to define where to store the model upon completion of the clustering. If a process is suspended and the model is stored, the user can also use this screen to browse the computer disk storage for different previously stored models.
Current data regarding the dataset CS is depicted in a panel 340 of the screen. Text boxes 342, 344, 346, 348 in this panel indicate a number of subclusters c, and average, minimum and maximum variances for the subclusters using the above definition of variance. A last text box 350 indicates the average size of the subclusters in terms of data points or tuples in the subclusters.
Additional command buttons allow the user to interact with the clustering as it occurs. A stop button 360 stops the clustering and stores the results to disk. A continue button 362 allows the process to be suspended and resumed by activating a resume button 364. A generate batch button 366 allows the user to generate a clustering batch file which can be executed as a separate process. Finally a close button 368 closes this window without stopping the clustering.
Multiple Model Embodiment In accordance with an alternate embodiment of the present invention, the process of Figure 4 is supplemented with a model optimizer. In accordance with this embodiment, a multiple number of different clustering models S are simultaneously generated by the computer 20. The multiple model clustering is depicted in the schematic of Figure 2 and the data structure of Figure 8. The disclosed procedures for data compression and storing of sufficient statistics allows updating multiple models
22 simultaneously, within a single data scan. K- eans, as well as many other members of the iterative clustering algorithms, are well-known to be extremely sensitive to initial starting condition. However, standard practice usually calls for trying multiple solutions from multiple random starting points. One process for choosing multiple starting points is disclosed in copending patent application entitled "A method for refining the initial conditions for clustering with applications to small and large database clustering." Which was filed in the United States Patent and Trademark Office on March 4, 1998 and is assigned to the assignee of the present invention.
To support standard practice in clustering, the invention includes the ability to explore multiple models. The key insights for this generalization are:
Retained points RS and the sets CS (representing local dense structures) are shared amongst the all models;
1 ) Each model, say Mi, will have its own discarded data sets DSM, (K sets, one for each cluster for each model) — if there are in models, there are a wxK discard sets; 2) The sufficient statistics for discarded data sets DSy for one of the models Mi are simply viewed as members of the global CS by all models other than Mi.
The overall architecture remains the same as the one shown in Figure 2, except that model updating and data compression steps are now performed over multiple models. Besides these observations there is at least one other data compression item worthy of further discussion: data discard order when multiple models are present. The algorithm decides on an individual data point basis which discard set fits it best. A data point that qualifies as a discard item for two models simply goes to the discard set of the model that it "fits" best. A data point cannot be allowed to enter more than one discard set else it will be accounted multiple times. Let x qualify as a discard item for both models M l and M2. If it were admitted to both, then model Ml will "feel" the effect of this point twice: once in its own DSJ and another time when it updates over DS2 which will be treated as part of CS as far as MI is concerned. Similarly for M2. By entering in exactly one discard set, say DSJ, the point x still affects M2 when M2 updates over CS and counts DSJ as part of CS.
There is an array S pointers mi ... ms where each pointer points to a different model data structure. The model data structure is depicted in Figure 6D. In this embodiment the structure CS and RS are shared by the multiple models. Each of the models ms is initialized with a different set of parameters (e.g. centroid vectors in case of K-means) (value of 'sum', M = 1) for the K different clusters of the model. When data is gathered at the step 1 10, that data is used to update each of the S models. An extended CLUSTERING procedure for the multiple model process takes into account the multiple model aspects of the structures DS and CS is perfromed on each of the S models. On a first iteration through the Figure 4 process there is no DS or CS dataset for any of the models so that all data is in the RS dataset. A given data point rs in the data set RS is compressed into the dataset DSjfor only one of the S models even though it may have a high probability of membership to another DS set. The data point rs is assigned to the set DS of the model that gives it the highest probability of membership (in case of K-means this would be the DS associated with cluster having centroid that is closest to the point rs).
DS structures for all the S models are determined by compressing data points into the appropriate DS data structure. The CS data structures for each of the models are then determined from the points remaining in RS. When performing the extended CLUSTERING procedure 120, however, the CS sufficient statistics must be augmented with the sufficient statistics contained in the DS data structures of the other models. When performing the extended CLUSTERING procedure to update a given model m,, the subclusters in CS must be augmented with the DS structures from the other models. Specifically, when updating model m,-, the extended K-means procedure considers the augmented set CSj = CS U (union) DSi U DS2 ... DSj.j U DSj+i U ...DS,when performing the loop 230 of Figure 7. If a data point is compressed into DS, it enters the DS set of only one model at the step 240, hence there is no double counting of data The multiple model analysis can be performed until one of the models satisfies the stopping criteria at the step 140. An alternate system would continue to compute all the models until each model reaches a stopping criteria. Additionally, the scalable K- means process could be performed until a certain percentage of the models have reached a stopping criteria. The multiple model implementation shares data structures between models and performs calculations on certain data unique to a given model. This analysis is susceptible to parallel processing on a computer 20 having multiple processing units 21
24 Thus far it is assumed that each of the data points that is accessed and used in the the K-means clustering procedure has the same weight as other data. Recall, however, that the process can allow suspension of a model build and resumption at a later time. The stopping criteria can be reached and the model stored and later used for whatever purpose the user desires. It is therefore possible that the database can be updated between clustering. In those instances it may be desirable to weight more recently added data more heavily than the data records used to cluster a database from a prior time. In this instance the recent records can be weighted by a heavier factor such as 1.5 per data point. Thus, when updating the Sum, Sumsq, and M components of the various data structures the weighting factor is applied to the data before updating the summaries.
By assigning the data to one of K clusters the invention provides a way to add an additional attribute or dimension to the data, i.e. cluster number. This additional attribute can be used as an index field for the data in the database. It can also be used as a pivoting attribute tυ reorganize and present the data in a different fashion to the user of the database.
Computer System With reference to Figure 1 an exemplary data processing system for practicing the disclosed data mining engine invention includes a general purpose computing device in the form of a conventional computer 20, including one or more processing units 21, a system memory 22, and a system bus 23 that couples various system components including the system memory to the processing unit 21. The system bus 23 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. The system memory includes read only memory (ROM) 24 and random access memory (RAM) 25. A basic input/output system 26 (BIOS), containing the basic routines that helps to transfer information between elements within the computer 20, such as during start-up, is stored in ROM 24.
The computer 20 further includes a hard disk drive 27 for reading from and writing to a hard disk, not shown, a magnetic disk drive 28 for reading from or writing to a removable magnetic disk 29, and an optical disk drive 30 for reading from or writing to a removable optical disk 3 1 such as a CD ROM or other optical media. The
25 hard disk drive 27, magnetic disk drive 28, and optical disk drive 30 are connected to the system bus 23 by a hard disk drive interface 32, a magnetic disk drive interface 33, and an optical drive interface 34, respectively. The drives and their associated computer-readable media provide nonvolatile storage of computer readable instructions, data structures, program modules and other data for the computer 20. Although the exemplary environment described herein employs a hard disk, a removable magnetic disk 29 and a removable optical disk 31 , it should be appreciated by those skilled in the art that other types of computer readable media which can store data that is accessible by a computer, such as magnetic cassettes, flash memory cards, digital video disks, Bernoulli cartridges, random access memories (RAMs), read only memories (ROM), and the like, may also be used in the exemplary operating environment.
A number of program modules may be stored on the hard disk, magnetic disk 29, optical disk 3 1, ROM 24 or RAM 25, including an operating system 35, one or more application programs 36, other program modules 37, and program data 38. A user may enter commands and information into the computer 20 through input devices such as a keyboard 40 and pointing device 42. Other input devices (not shown) may include a microphone, joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to the processing unit 21 through a serial port interface 46 that is coupled to the system bus, but may be connected by other interfaces, such as a parallel port, game port or a universal serial bus (USB). A monitor 47 or other type of display device is also connected to the system bus 23 via an interface, such as a video adapter 48. In addition to the monitor, personal computers typically include other peripheral output devices (not shown), such as speakers and printers.
The computer 20 may operate in a networked environment using logical connections to one or more remote computers, such as a remote computer 49. The remote computer 49 may be another personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the computer 20, although only a memory storage device 50 has been illustrated in Figure 1. The logical connections depicted in Figure 1 include a local area network (LAN) 51 and a wide area network (WAN) 52.
26 Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.
When used in a LAN networking environment, the computer 20 is connected to the local network 51 through a network interface or adapter 53. When used in a WAN networking environment, the computer 20 typically includes a modem 54 or other means for establishing communications over the wide area network 52, such as the Internet. The modem 54, which may be internal or external, is connected to the system bus 23 via the serial port interface 46. In a networked environment, program modules depicted relative to the computer 20, or portions thereof, may be stored in the remote memory storage device. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.
While the present invention has been described with a degree of particularity, it is the intent that the invention include all modifications and alterations from the disclosed implementations falling within the spirit or scope of the appended claims.
The following APPENDICES describe components of the scalable clustering analysis Appendix A - Worst Case Analysis
Assume that the following functions are available:
1. [tValue] = ComputeT( alpha , DegreesFreedom): computes the t-statistics for the given value of alpha (1 -ConfidenceLevel, in our case) and DegreesFreedom (# of points in a cluster -1 in our case). 2. DSCopy( DSNew, DSOrig): copies DSOrig into DSNew, DSNew is altered, DSOrig remains the same.
3. [Weight Vec] = ComputeWeightVec( DataPoint, Model ): for K-Means, returns the {0, 1 } weight vector indicating the cluster to which DataPoint is assigned.
4. AddRS( DataPoint, RS ): appends singleton DataPoint to the end of RS, RS is increased by 1 point.
The following functions determine the new sets DS and RS from the old sets, the current Model, and given ConfidenceLevel which is a number between 0 and 1. [ DSNew, RSNew ] = WorstCaseAnalysis( DS, RS, Model, ConfidenceLevel )
{
RSNew = empty; k = Length( Model ); r = Length( RS );
27 // first determine values for LOWER and UPPER alpha = 1 -ConfidenceLevel; for l = l,...,k
{ Mean = Model(l).GetMean();
CVDiag = Model(l).GetCVDiag();
TValue = ComputeT( (alpha/n), Model(l).GetNum()-l ); // correct t- value
For j = l,...,n {
LOWER(l).LowVec(j) = MeanQ) - (TValue)* sqrt( CVDiag(j)/(Model(l).GetNum())); UPPER(l).UpVec(j) = Mean(j) + (TValue)*
Sqrt( CVDiag(j)/(Model(l).GetNum))); }
}
// Copy DS into DSNew DSCopy( DSNew, DS );
// for each singleton element in r, perform the worst-case "jiggle" and determine
// if the cluster assignment for this data point changes, if so keep in RS, if not, put
// in DSNew forj = l,...,r {
DataPoint = RS(j).RSVec;
// Determine the cluster to which this data point is assigned
[TrueWeightVec] = ComputeWeightVec( DataPoint, Model );
// Zero out the perturbed model Zero( PerturbModel );
// determine the perturbed model for this data point for l - l ,...,k
{
Mean = Model(l).GetMean(); If ( TrueWeighVec(l) == 1.0) // DataPoint is assigned to cluster
1
{
// the pertubed model center is as far away from DataPoint // as possible for h = l, ...,n
{ if ( DataPoint(h) >= Mean(h) )
{ PerturbModel(l).Center(h) =
LOWER(l).LowVec(h);
} else
28 {
PerturbModel(l).Center(h) = UPPER(l).UpVec(h);
} }
} else
{
// the data point is not assigned to model 1, move the // perturbed
// center as close to data point as possible for h = l, ... ,n
{ case ( DataPoint(h) >= UPPER(l).UpVec(h) )
PerturbModel(l).Center(h) = UPPER(l).UpVec(h);
} case ( DataPoint(h) <= LOWER(l).LowVec(h) ) {
PerturbModel(l).Center(h) =
LOWER(l).LowVec(h);
} otherwise {
PertubModel(l).Center(h) = DataPoint(h);
// at this point the perturbed model has been determined for the given data // point // determine the assignement of this point under the perturbed model
[PerturbWeightVec] = ComputeWeightVec( DataPoint, PerturbModel
);
// determine if the assignments are the same. If so, update the correct // DSNew
// component. If not, put the point in RSNew
for l = 1 k
{ if ( TrueWeightVec(l) == 1 .0) and ( PerturbWeightVec(l) ==
1.0)
{
DSNew(l).Sum += DataPoint;
2 DSNew(l) Num ++,
if ((TrueWeightVec(l) = 1 0) and ( Perturb Weight Vec(l) = 0 0)) or ((TrueWeightVec(l) == 0 0) and
( PerturbWeightVec == 1 0))
{
AddRS( DataPoint, RSNew ),
} }
} return DSNew, RSNew,
} [ DSsNew, RSNew ] = WorstCaseAnalysis( DSs, RS, Model, ConfidenceLevel )
{ for (1 = 1 , I <= Length(DSs), 1++) { [DSsNew[l], RSNew] =
WorstCaseAnalysis(DSs[I],RS,Models[I], ConfidenceLevel); return DSsNew,RSNew,
Appendix B-Mahalanobis Threshold Analysis
Assume that the following functions are available
1 [WeightVec] = ComputeWeightVec( DataPoint, Model), in the k-Mean case, returns the {0, 1 } weight vector indicating the cluster to which DataPoint is assigned
2 AddRS( DataPoint, RS) appends singleton DataPoint to the end of RS, RS is increased by 1 point
3 RSDistSort( RSDist ) sorts the list RSDistSort from smallest to greatest by the values in MahalDist 4 [Dist] = DistanceMahalanobis( DataPoint, Center, CVDiagonal ) computes the
Mahalanobis distance from DataPoint to Center with the given covariance diagonal
CVDiagonal 5 [integer] = Floor(float) returns the integer obtained by rounding the float down to the nearest integer 6 DSCopy(DSNew, DSOrig) copies DSOrig into DSNew, DSNew is altered,
DSOrig remains the same
The method
[ DSsNew, RSNew ] = Threshold( DSs, RS, Models, Percentage )
{
// Percentage is the percentage of points to compress with this function
30 DSCopy( DSNew, DS ), RSNew = empty,
Initialize(RSDist), // initialize the RSDist structure k = Length( Model ), r = Length( RS ),
// Fill the fields of the RSDist structure For 1 = 1, ,K
{
Let Model = Models(I), Forj = 1 , . ,r
{
DataPoint = RS0).RSVec,
[Weight Vec] = ComputeWeightVec( DataPoint, Model ); for l = 1 ,. .,k
{ if ( Weight Vec(l) == 1 0 )
I
// DataPoint is assigned to cluster 1 RSDistO) Modellndex = I,
RSDist(j) RSIndex = j, RSDistO) ClustAssign = 1, RSDistO) MahalDist = DistanceMahalanobis (DataPoint, Model(l).GetMean(),
Model(l) GetCVDiag() ),
}
RSDistSort( RSDist ), // do the sorting
// determine the number of points to compress CompressNum = Floor( r*Percentage ),
Forj = 1,. .,r
{
DataPoint = RS(RSDistO) RSIndex) RSVec, if 0 <- CompressNum) {
DSsNew(RSDistO) Modellndex,RSDistO) ClustAssign ) Sum += DataPoint,
DSsNew(RSDistO) ModelIndex,RSDistO).ClustAssign)SumSq += DataPoint*DataPoint, DSsNew(RSDistO) Modellndex, RSDistO) ClustAssign ).Num
++,
} else {
AddRS( DataPoint, RSNew );
}
} return DSsNew, RSNew;
Appendix C SubCluster data set CS We assume that the following functions are availale:
1. [Model] = VanillaKMean( Model , Data , StopTol ): takes initial values for Model and updates the model with the Data until the model ceases to changes, within StopTol. The updated model is returned in Model.
2. [CSMerged] = MergeCS( CSEle l, CSElem2 ): take the sufficient statistics for 2 sub-clusters, CS l and CS2, and computes the sufficient statistics for the subcluster formed by merging CS l and CS2.
3. AddRS( DataPoint, RS): appends singleton DataPoint to the end of RS, RS is increased by 1 point.
4. [SubModel] = RandomSample( RSNew, kPrime ): randomly chooses kPrime elements from RSNew and uses these as the initial points for the vanilla k-mean algorithm. The elements are stored in SubModel.
5. [Weight Vec] = ComputeWeightVec( DataPoint, Model ): computes the {0,1 } weight vector with k elements. DataPoint is assigned to cluster j if the j-th element of the Weight Vec is 1 , the other elements are 0. 6. Append( integer, integerList ): appends the integer to the end of integerList.
7. RemoveCSCandidates( IndexList, CS ): removes the elements specified in IndexList from CS and returns the altered, smaller list in CS.
8. [BigCSList] = AppendCS( CSList l, CSList2 ): creates the BigCSList by appending CSList2 to the end of CSList l . 9. [SubcPartner,SubcPartnerInd] = FindMergePartner( Index, CS ): finds the element (subcluster) in CS that has center nearest to CS(Index) (a different subcluster) and returns this element in SubcPartner and its index.
10. AppendCSEnd( CS. CSElement ): appends CSElement to the end of the CS list
1 1. [CSlndex] = FindCSMatch(DataPoint,CS): finds the cluster in CS to which DataPoint is closest and, if DataPoint was merged into that cluster, the "density" criteria would still be satisfied. If no such cluster exists, then this routine returns NULL.
12. [CSMergedElem] = MergeCS(DataPoint,CSElem): merges a singleton data point into a CS cluster and returns the merged Cluster.
The subclusterin og method:
[CSNew, RSNewer] = CSUpdate( CS, RSNew, StdTol, PointsPerSubClust, MinPoints, StopTol ) {
// STDTol is a scalar which defines the "dense" criteria discussed above
// a subcluster is deemed "dense" if the square root of the maximum element of
// its covariance matrix is less than StdTol // PointPerSubClust is an integer which used to determine the number of
// secondary
// subclusters to search for in RSNew The number of sub-cluster to search for is // (# of points in RSNew)/(PointsPerSubClust)
// MinPoints is an integer specifying the minimum number of points allowed in a
// subcluster If a sub-cluster does not have have this number of points, the points // remain as singletons are are placed in RSNewer
// StopTol is a tolerance for the vanilla k-mean algorithm
//prefiltering
//filter as many points from rs into cs as possible now r = length(RSNew), for i = 1, ,r
{
DataPoint = RSNew(i) RSVec,
[Weight Vec] = computeWeightVec(DataPoint,CS), for j = 1, , CS NumElems
if (Weight VecO) == 1 0) { if (OkToMerge(DataPoint,CS0))){
CSO) = Merge(DataPoint,CSO)), RSNew Remove(i),
}
CSCopy( CSNew, CS), // copy CS into CSNew RSNewer = empty,
// determine the number of singleton data elements r = length( RSNew ),
// kPrime = the number of "dense" regions in RSNew to search for kPrime = floor( r/ PointsPerSubClust ),
// choose the starting point for the vanilla k-Mean algorithm as kPrime random
// elements // of RSNew [SubModel] = RandomSample( RSNew, kPrime ), // cluster these points and return the sub-clusters in CSCandidates
[CSCandidates] = VanillaKMean( SubModel, RSNew, StopTol ),
// Filter out the candidates that have fewer than MinPoints and put the points // generating
// these candidates into RSNewer CSCandidateRemoveList = empty; for 1 = 1, ... , kPrime { if (CSCandidates(l).Num < MinPoints)
{
// add the points in RSNew nearest to this sub-cluster center to RSNewer forj = 1 r
{
DataPoint = RSNewO).RSVec; [Weight Vec] = ComputeWeightVec( DataPoint, CSCandidates ); if( Weight Vec(l) == 1.0)
{
// this data point is in this sub-cluster AddRS( RSNewer, DataPoint ); Appendø, CSCandidateRemoveList);
}
}
// remove those candidates not having enough points in them RemoveCS( CSCandidateRemoveList, CSCandidates );
kDoublePrime = length( CSCandidates ); CSCandidateRemoveList = empty;
// filter out the candidates that do not satisfy the "dense" criteria for 1 = 1 , . .. , kDoublePrime
{
CVDiag = CSCandidates(l).GetCVDiag();
// note that in the next line, sqrt is applied to all of the n elements of // CVDiag and
// then the max is over this list of n elements If (max(sqrt(CVDiag)) > StdTol)
{
// this sub-cluster's max standard deviation is too large Appendø, CSCandidateRemoveList );
} }
// remove those candidates satisfying this "dense" criteria RemoveCS( CSCandidateRemoveList, CSCandidates ); [CSNew] = AppendCS(CS, CSCandidates);
// now consider merging the elements of CSNew done = false;
34 CSIndex = I N/ start with the first element of CSΝew while (not done)
{
// find the partner for CSIndex element of CSΝew [CSMergePartner,CSMergePartnerIndex] = FindMergePartner(CSIndex,CSΝew);
// merge the two
[CandMerge] = MergeCS( CSMergePartner , CSNew( CSIndex ) );
// see if the merged sub-cluster still satisfies "density" criterion if ( max( sqrt( CandMerge GetCVDiagO )) < StdTol )
{
// the merged cluster is "dense" enough
// remove CSNew(CSIndex) and CSMergePartner from list CSNewRemoveList = [ CSIndex, CSMergePartnerlndex ]; RemoveCS(CSNewRemoveList, CSNew), // append the merged cluster to the end of CSNew
AppendCSEnd(CSNew, CandMerge);
// notice by doing the remove, the next sub-cluster to use // consider merging is CSNew( CSIndex ) so there is no need to // increment CSIndex
} else
{
// the merged cluster is not "dense" enough // do nothing and increment CSIndex
CSIndex++,
// See if we've considered mergin all the sub-clusters if (CSIndex > length( CSNew ))
{ done = true;
} } return CSNew, RSNewer; }
Appendix D stopping criteria Method 1
[Stop] = StoppingCritU OldModel, NewModel, PrevModelDiff, StopTol ) {
3 s // OldModel holds the model parameters (means, etc.) over calculated over the
// last iteration
// NewModel holds the model parameters over calculated over the current
// iteration
// PrevModelDiff is a vector of model deltas over the last (current-
1 ),..., (current
II z)
II iterations
// StopTol is a scalar stopping tolerance k = length( OldModel );
// determine the model difference between OldModel and NewModel NewDif = 0.0; for 1 = l,...,k
{
OldMean = OldModel(l).GetMean(); NewMean = NewModel(l).GetMean(); NewDiff += Distance2Norm( OldMean , NewMean ); }
NewDiff = (NewDiff/k);
If ( max( [PrevModelDiff , NewDiff] ) < StopTol )
{ Stop = true;
} else
{
Stop = false; } return Stop;
Method 2:
Stopping criteria 2 requires the following function which computes the "Energy" of a given Model given the sets DS, CS and RS: [Energy] = EnergyFunctionKMean( Model, DS, CS, RS )
{ k = length( Model ); c = length( CS ); r = length( RS );
Energy = 0.0;
// compute energy over the set RS
36 forj = 1 , ,r
{
DataPoint - RSø) RSVec,
[Weight Vec] = ComputeWeightVec( DataPoint, Model ), for 1 = 1 , ,k
{ if (Weight Vec(l) = 1 0)
{
Energy += Distance2Norm(DataPoιnt,Model(l) GetMean()),
}
}
} // compute energy over the set CS
CSPoints = 0, // count the number of data points summarized by CS Forj = 1, ,c
{
CSCenter = CSø) GetMean(), CSPoints += CS(j) GetNum,
[WeightVec] = ComputeWeightVec( CSCenter, Model ), for 1 = 1, ,k
{ if ( Weight Vec(l) == 1 0) {
Energy += CSø) GetNum()*
Dιstance2Norm(CSCenter, Model(l) GetMean()),
} }
// compute the energy over DS
DSPoints = 0, // count the number of points summarized by DS For 1 = 1 , ,k
{
DSCenter = DS(1) GetMean(), DSPoints += DS0) GetNum(), Energy += DS0) GetNum*Distance2Norm(DSCenter, Model(l) GetMean()),
}
Energy = ( l/(r+CSNum+DSNum))*Energy,
Return Energy,
}
The method
[Stop] = StoppingCrit2( OldModel, NewModel, DSOId, DSNew, CSOld, CSNew,
17 RSOld, RSNew, PrevEnergyDiff, StopTol )
{
// OldModel holds the model parameters (means, etc.) calculated over the // last iteration // NewModel holds the model parameters over calculated over the current
// iteration
// PrevEnergyDiff is a vector of energy deltas over the last (current- 1 ),..., (current-z) // iterations // StopTol is a scalar stopping tolerance
// determine the difference in energy between the new and old models NewEnerg = EnergyFunctionKMean( NewModel, DSNew, CSNew, RSNew OldEnergy = EnergyFunctionKMean( OldModel, DSOId, CSOld, RSOld );
NewDiff = abs(NewEnergy-OldEnergy),
If ( max( [PrevModelDiff , NewDiff] ) < StopTol )
{
Stop = = true,
} else
{
Stop = = false.
}
Figure imgf000040_0001
return Stop,
Appendix E Multiple Model Updating The following methods are assumed to be available to the clustering driver'
1 UpdateModel (Model,RSNode,WeightVec) updates Model with singleton data point RSNode using weighting information specified in Weight Vec.
2 UpdateModel (Model, CSNode,WeightVec) updates Model with compressed set node CSNode using weighting information specified in Weight Vec
3 UpdateModel(Model,DSNode,WeightVec) updates Model with discard set node using weighting information specified in WeightVec
4 ReinitializeEmptyCluster(ModelNode,DSs,CS,RS) reinitializes an empty cluster ModelNode, using information in DSs, CS and RS Currently empty clusters are re-initialized using the RS and CS nodes which are most unprobable w.r.t. to their most probable clusters However, many other heuristics are possible.
The method [ModelsNew] = Update_Models (Models,DSs,CS,RS,stopTol)
{ r = Length(RS), c = Length(CS), or (iModel = 1 ; iModel < Length(Models); iModel++) {
// Determine number of clusters. Let Model = Models[iModel];
K = Length(Model);
// Perform Extended Update algorithm. do { // Initialize ModelTemp to zeros.
Zero(ModelTemp);
// Update ModelTemp over RS. for (i = 1 ; i <= r; i++) { [WeightVec] = computeWeightVec(RS(i).RSVec,Model);
UpdateModel(ModelTemp,RS(i), WeightVec);
// Update ModelTemp over CS. for (i = 1 ; i <= r; i++) {
[WeightVec] = computeWeightVec(CS(i).GetMean()); UpdateModeløvlodelTemp, WeightVec);
}
// Update ModelTemp over DSs. for (i = 1 ; i <= Length(DS); i++) { Let DS = DSsø) for 0 = 1; j <= Length(DS); j++) { [WeightVec] = computeWeightVec(DSO),Model);
UpdateModel(ModelTemp,DSø), WeightVec);
// Re-initialize ModelTemp if (ReinitializeEmptyClusters) { for (i = 1; i <= K; i++) { if (ModelTemp(i).GetCountO == 0) {
ReinitializeEmptyCluster(ModelTemp(i),DSs ,CS,RS);
}
[Diff] = ModelDiff(Model,ModelTemp);
} while (Diff > StopTol); [Model] = ModelCopy(ModelTemp);
9 } return [Models]
}
Appendix E Scaleable Clustering Driver Method
The following methods are assumed to be available to the clustering driver:
1. [dataSource] = Open_DataSource(dataSourceName,scanType): opens up a connection to the given data source, from which all data points will be consumed. 2. [RS,numRowsRead] = DataSource_jgetNextRows(dataSource,maxRSRows): retrieves up to maxRSRows from dataSource, using the appropriate scan type.
3. [ModelsNew] = ModelsUpdate(Models,DSs,CS,RS,stopTol): see above description of ModelsUpdate.
4. [DSsNew, RSNew] = DSsUpdate(DSs,RS,ModelsNew,Confidence): performs primary data compression on RS, see above description.
5. [CSNew,RSNewer] = CSUpdate(CS,RSNew,StdTol, PointsPerSubClust, MinPoint, StopTol): performs secondary compression on RS, see above description.
6. [Stop] = StopCriteria (Models,DSs,CS,RS,OldEnergy,PrevDiff ModelsNew,DSsNew,CSNew,RSNewer,StopTol): performs stop criteria check, see above description.
7. Close_DataSource(dataSource): closes connection to data source.
The method:
[Models] = Cluster ()
{
// Initialize all of the models ... for (iModel = 1; iModel < numModels; iModel++) { InitializeModel (Models[iModel]);
Zero(DSs[iModel]); PrevDiffsfiModel] = 0; Stop[iModel] = 0; OldEnergy[iModel] = 0; }
// Connect to the data source ...
[dataSource] = Open_DataSource(dataSourceName); do {
// Fill up RS to specified maximum level.
[RS,numRowsRead] = datasrc_getNextRows(dataSource,maxRSRows); if (numRowsRead == 0) break;
// Update Models, DSs, CS and RS.
[ModelsNew] = ModelsUpdate(Models,DSs,CS,RS,stopTol);
40 [DSsNew,RSNew] = DSUpdate(DSs,RS,ModelsNew,Confidence); [CSNew,RSNewer] =
CSUpdate(CS,RSNew,StdTol,PointsPerSubClust,MinPoint,Sto pTol);
// Perform Stop Criteria check.
[Stop] = StopCriteria (Models,DSs,CS,RS,OldEnergy,PrevDiff,
ModelsNew,DSsNew,CSNew,RSNewer,StopTol); // Copy new values.
[Models] = [ModelsNew];
[DSs] = [DSsNew];
[CS] = [CSNew];
[RS] = [RSNewer]; } while (! Stop);
Close_DataSource(dataSource); return [Models];
41

Claims

Claims 1. A method for clustering data in a database that is stored on a storage medium comprising the steps of: a) obtaining a portion of the data in the database from a storage medium; b) clustering data from the portion of data obtained from the database based upon any clustering criteria; c) summarizing at least some of the data contained within the portion of data based upon a compression criteria to produce sufficient statistics for the data satisfying the compression criteria; and d) continuing to obtain portions of data from the database and characterizing the clustering of data in the database from newly sampled data and the sufficient statistics until a specified criteria has been satisfied.
2. The method of claim 1 wherein portions of data obtained from the database are stored in a rapid access memory smaller in size than the storage requirements of the database and wherein the sufficient statistics replace data obtained from the database that was stored in the rapid access memory to allow further sampling of data from the database to be brought into the rapid access memory for clustering.
3. The method of claim 1 wherein the sufficient statistics are stored in a rapid access memory and the step of characterizing the clustering of data in the database utilizes both stored sufficient statistics and newly obtained data from the database.
4. The method of claim 1 wherein the sufficient statistics are stored in a rapid access memory and the step of characterizing clustering of the database utilizes stored sufficient statistics, non-summarized data from an earlier data acquisition step and newly obtained data from the database.
5. The method of claim 1 wherein the specified criteria stops further characterization based on a comparison of different database models derived from data obtained from
42 the database and the characterization is stopped when a change in said models is less than a specified stopping criteria.
6. The method of claim 1 wherein the specified criteria suspends a construction of a clustering model to allow the construction to be resumed at a later time.
7. The method of claim 6 wherein the clustering of data is performed by a computer and wherein the database is updated with additional data before completion of the construction of a clustering model.
8. The method of claim 1 additionally comprising the step of displaying progress of the characterization of the clustering of the database on a user interface and provides a user controller input for stopping or suspending further building of a database clustering model.
9. The method of claim 8 wherein the displaying step comprises the step of displaying a user interface for allowing a user to suspend and resume the clustering.
10. The method of claim 8 additionally comprising the step of providing a user interface wherein the user can adjust parameters of the clustering analysis.
11. The method of claim 10 wherein the parameters include a stopping criteria.
12. The method of claim 1 wherein the specified criteria corresponds to receipt of a suspend command to suspend the clustering while the database is updated before again resuming the clustering.
13. The method of claim 1 additionally comprising the step of preparing a clustering model and using the model to prepare a histogram, summary, of data as an input to a visualization system.
43
14. The method of claim 1 wherein the data is obtained from the database on a random, indexed, or sequential basis.
15. The method of claim 1 wherein the data is initially obtained from the database on a random basis and after at least one iteration is switched to a sequential or indexed basis.
16. The method of claim 1 wherein the step of summarizing the data includes the substep of classifying the data into at least two groups of data and wherein data in a first group is compressed and data in a second group is maintained as individual data records of the database.
17. The method of claim 16 wherein two subgroups of data in the first group are summarized and wherein the data in one of the subgroups is classified into one or more subclusters having sufficient statistics for each subcluster.
18. The method of claim 17 wherein a clustering model is updated each time a portion of the data is accessed from the database and wherein data read into the rapid access memory from the database is added to one of the subclusters before forming new subclusters from the data accessed from the database.
19. The method of claim 17 wherein data that is accessed from the database is stored as data vectors represented as either fixed length records or varying length records and wherein the data vectors are combined with existing subclusters before using the data vectors to form new subclusters.
20. A computer readable medium having stored thereon a data structure, comprising: a) a first data portion containing a model representation of data stored on a database; b) a second data portion containing sufficient statistics of a portion of the data in the database; and
44 c) a third data portion containing individual data records obtained from the database for use with the sufficient statistics to determine said model representation contained in the first data portion.
21. The data structure of claim 20 wherein the second data portion includes a first subportion that is organized in accordance with data clusters which characterize the database and a second subportion which is not organized according to cluster.
22. The data structure of claim 20 wherein the data associated with a data cluster is a vector containing a summation of data records from the database wherein the vector components are attributes of said data records and further wherein the data associated with a data cluster includes the number of data records from the database associated with said cluster.
23. In a computer data mining system, apparatus for evaluating data in a database comprising the steps of a) one or more data storage devices for storing a database of records on a storage medium; b) a computer having an interface to the storage devices for reading data from the storage medium and bring the data into a rapid access memory for subsequent evaluation; and c) said computer comprising a processing unit for evaluating at least some of the data in the database and for characterizing the data into multiple numbers of data clusters; said processing unit programmed to retrieve a subset of data from the database into the rapid access memory, evaluate the subset of data to further characterize the database clustering using a clustering criteria, and produce a summarization of at least some of the retrieved data before retrieving additional data from the database.
24. The apparatus of claim 23 wherein the processing unit further comprises means to iteratively bring data from the database and update the characterization of data into clusters until a specified criteria has been reached.
45
25. The apparatus of claim 23 additionally comprising a video monitor to depict a user interface which updates a user regarding a status of the classification of data and including a command input for allowing a user to suspend or to stop the process.
26. The apparatus of claim 25 wherein the computer presents a user interface on the video monitor for allowing a user to adjust parameters of the clustering process .
27. The apparatus of claim 26 wherein the user interface allows a user to adjust a depiction or organization of data from the database based upon a cluster number to which data is assigned during a clustering process.
28. The apparatus of claim 23 wherein the computer updates multiple clustering models and includes multiple processors for updating said multiple clustering models.
29. In a computer data mining system, a method for evaluating data in a database that is stored on a storage medium comprising the steps of: a) initializing multiple storage areas for storing multiple cluster models of the data in the database; b) obtaining a portion of the data in the database from a storage medium; c) using a clustering criteria to characterize a clustering of data from the portion of data obtained from the database for each model; c) summarizing at least some of the data contained within the portion of data based upon a compression criteria to produce sufficient statistics for the data satisfying the compression criteria; and d) continuing to obtain portions of data from the database and characterizing the clustering of data in the database from newly sampled data and the sufficient statistics for each of the multiple cluster models until a specified criteria has been satisfied.
30. The method of claim 29 wherein a portion of the sufficient statistics is unique for each of the clustering models and wherein a portion of the sufficient statistics is shared between different clustering models.
46
31. The method of claim 29 wherein the specified criteria is reached when iterative solutions for one of the models does not vary by more that a predetermined amount.
32. The method of claim 29 wherein the specified criteria is reached when iterative solutions for a specified number of the models do not vary by more that a predetermined amount.
33. The method of claim 29 wherein the multiple models are generating from multiple different starting conditions.
34. The method of claim 33 wherein at least some of the models have different numbers of clusters.
35. The method of claim 29 wherein multiple clustering models are generated in an inner processing loop and wherein an outer processing loop iterates over different values of cluster number.
36. A method for clustering data in a database that is stored on a storage medium comprising the steps of: a) obtaining a portion of the data in the database from a storage medium; b) clustering data from the portion of data obtained from the database based upon any clustering criteria; c) summarizing at least some of the data contained within the portion of data based upon a compression criteria to produce sufficient statistics for the data satisfying the compression criteria; and d) obtaining other data portions from the database and characterizing the clustering of data in the database from newly sampled data and the sufficient statistics until a specified criteria has been satisfied; e) wherein data sampled from the database in assigned different weights for defining a clustering model based upon the data portion.
47
PCT/US1999/005759 1998-03-17 1999-03-16 A scalable system for clustering of large databases WO1999048018A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP99913907A EP1062590B1 (en) 1998-03-17 1999-03-16 A scalable system for clustering of large databases
DE69938339T DE69938339T2 (en) 1998-03-17 1999-03-16 A SCALABLE SYSTEM FOR GROUPING LARGE DATA BENCHES

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US09/040,219 1998-03-17
US09/040,219 US6374251B1 (en) 1998-03-17 1998-03-17 Scalable system for clustering of large databases

Publications (1)

Publication Number Publication Date
WO1999048018A1 true WO1999048018A1 (en) 1999-09-23

Family

ID=21909787

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US1999/005759 WO1999048018A1 (en) 1998-03-17 1999-03-16 A scalable system for clustering of large databases

Country Status (5)

Country Link
US (1) US6374251B1 (en)
EP (1) EP1062590B1 (en)
AT (1) ATE389213T1 (en)
DE (1) DE69938339T2 (en)
WO (1) WO1999048018A1 (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008034213A1 (en) * 2006-09-18 2008-03-27 Infobright Inc. A method and system for data compression in a relational database
WO2011056099A1 (en) 2009-11-03 2011-05-12 Telefonaktiebolaget L M Ericsson (Publ) Reducing computational complexity during user data analysis
US8266147B2 (en) 2006-09-18 2012-09-11 Infobright, Inc. Methods and systems for database organization
US8417727B2 (en) 2010-06-14 2013-04-09 Infobright Inc. System and method for storing data in a relational database
US8521748B2 (en) 2010-06-14 2013-08-27 Infobright Inc. System and method for managing metadata in a relational database

Families Citing this family (122)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6581058B1 (en) * 1998-05-22 2003-06-17 Microsoft Corporation Scalable system for clustering of large databases having mixed data attributes
GB9903697D0 (en) * 1999-02-19 1999-04-14 Pattern Computing Limited A computer-based method for matching patterns
US6549907B1 (en) * 1999-04-22 2003-04-15 Microsoft Corporation Multi-dimensional database and data cube compression for aggregate query support on numeric dimensions
US6957189B2 (en) * 1999-08-30 2005-10-18 Sabre Inc. Apparatus and method for creating a marketing initiative
US8935198B1 (en) * 1999-09-08 2015-01-13 C4Cast.Com, Inc. Analysis and prediction of data using clusterization
US7424439B1 (en) * 1999-09-22 2008-09-09 Microsoft Corporation Data mining for managing marketing resources
US7106329B1 (en) 1999-09-30 2006-09-12 Battelle Memorial Institute Methods and apparatus for displaying disparate types of information using an interactive surface map
US6990238B1 (en) * 1999-09-30 2006-01-24 Battelle Memorial Institute Data processing, analysis, and visualization system for use with disparate data types
CN1241135C (en) * 1999-10-21 2006-02-08 国际商业机器公司 System and method of sequencing and classifying attributes for better visible of multidimentional data
US6643645B1 (en) * 2000-02-08 2003-11-04 Microsoft Corporation Retrofitting recommender system for achieving predetermined performance requirements
US7162482B1 (en) * 2000-05-03 2007-01-09 Musicmatch, Inc. Information retrieval engine
US6633882B1 (en) * 2000-06-29 2003-10-14 Microsoft Corporation Multi-dimensional database record compression utilizing optimized cluster models
US6944607B1 (en) * 2000-10-04 2005-09-13 Hewlett-Packard Development Compnay, L.P. Aggregated clustering method and system
US6584433B1 (en) * 2000-10-04 2003-06-24 Hewlett-Packard Development Company Lp Harmonic average based clustering method and system
US6922660B2 (en) * 2000-12-01 2005-07-26 Microsoft Corporation Determining near-optimal block size for incremental-type expectation maximization (EM) algorithms
US6519591B1 (en) * 2000-12-22 2003-02-11 Ncr Corporation Vertical implementation of expectation-maximization algorithm in SQL for performing clustering in very large databases
US6615205B1 (en) * 2000-12-22 2003-09-02 Paul M. Cereghini Horizontal implementation of expectation-maximization algorithm in SQL for performing clustering in very large databases
US6658423B1 (en) * 2001-01-24 2003-12-02 Google, Inc. Detecting duplicate and near-duplicate files
US6928434B1 (en) * 2001-01-31 2005-08-09 Rosetta Marketing Strategies Group Method and system for clustering optimization and applications
US7398270B1 (en) * 2001-01-31 2008-07-08 Choi Lawrence J Method and system for clustering optimization and applications
US6745184B1 (en) * 2001-01-31 2004-06-01 Rosetta Marketing Strategies Group Method and system for clustering optimization and applications
US6785684B2 (en) * 2001-03-27 2004-08-31 International Business Machines Corporation Apparatus and method for determining clustering factor in a database using block level sampling
US7155668B2 (en) * 2001-04-19 2006-12-26 International Business Machines Corporation Method and system for identifying relationships between text documents and structured variables pertaining to the text documents
US20030028504A1 (en) * 2001-05-08 2003-02-06 Burgoon David A. Method and system for isolating features of defined clusters
US7272617B1 (en) * 2001-11-30 2007-09-18 Ncr Corp. Analytic data set creation for modeling in a customer relationship management system
US7174343B2 (en) * 2002-05-10 2007-02-06 Oracle International Corporation In-database clustering
US7747624B2 (en) * 2002-05-10 2010-06-29 Oracle International Corporation Data summarization
US7133811B2 (en) * 2002-10-15 2006-11-07 Microsoft Corporation Staged mixture modeling
US20040093261A1 (en) * 2002-11-08 2004-05-13 Vivek Jain Automatic validation of survey results
US6993516B2 (en) * 2002-12-26 2006-01-31 International Business Machines Corporation Efficient sampling of a relational database
US20050033723A1 (en) * 2003-08-08 2005-02-10 Selby David A. Method, system, and computer program product for sorting data
US7539690B2 (en) * 2003-10-27 2009-05-26 Hewlett-Packard Development Company, L.P. Data mining method and system using regression clustering
US7403640B2 (en) * 2003-10-27 2008-07-22 Hewlett-Packard Development Company, L.P. System and method for employing an object-oriented motion detector to capture images
US7225200B2 (en) 2004-04-14 2007-05-29 Microsoft Corporation Automatic data perspective generation for a target variable
US7349914B1 (en) 2004-05-04 2008-03-25 Ncr Corp. Method and apparatus to cluster binary data transactions
US8078559B2 (en) 2004-06-30 2011-12-13 Northrop Grumman Systems Corporation System and method for the automated discovery of unknown unknowns
US8631347B2 (en) * 2004-11-15 2014-01-14 Microsoft Corporation Electronic document style matrix
US7415487B2 (en) * 2004-12-17 2008-08-19 Amazon Technologies, Inc. Apparatus and method for data warehousing
GB2422081B (en) * 2005-01-05 2010-09-29 Ling Dynamic Systems Statistical streaming
US7359913B1 (en) 2005-05-13 2008-04-15 Ncr Corp. K-means clustering using structured query language (SQL) statements and sufficient statistics
US7739314B2 (en) * 2005-08-15 2010-06-15 Google Inc. Scalable user clustering based on set similarity
US20070055708A1 (en) * 2005-09-07 2007-03-08 Ncr Corporation Processing formulae in rules for profitability calculations for financial processing in a relational database management system
US7840774B2 (en) * 2005-09-09 2010-11-23 International Business Machines Corporation Compressibility checking avoidance
US7721205B2 (en) * 2005-09-15 2010-05-18 Microsoft Corporation Integration of composite objects in host applications
US20070061351A1 (en) * 2005-09-13 2007-03-15 Microsoft Corporation Shape object text
US7783971B2 (en) * 2005-09-13 2010-08-24 Microsoft Corporation Graphic object themes
US20070061349A1 (en) * 2005-09-15 2007-03-15 Microsoft Corporation Hierarchically describing shapes
US8001526B2 (en) * 2005-09-15 2011-08-16 Microsoft Corporation Hierarchical property storage
US7716169B2 (en) * 2005-12-08 2010-05-11 Electronics And Telecommunications Research Institute System for and method of extracting and clustering information
US20070250476A1 (en) * 2006-04-21 2007-10-25 Lockheed Martin Corporation Approximate nearest neighbor search in metric space
US20070255684A1 (en) * 2006-04-29 2007-11-01 Yahoo! Inc. System and method using flat clustering for evolutionary clustering of sequential data sets
US8655623B2 (en) * 2007-02-13 2014-02-18 International Business Machines Corporation Diagnostic system and method
US7636715B2 (en) * 2007-03-23 2009-12-22 Microsoft Corporation Method for fast large scale data mining using logistic regression
US8402027B1 (en) * 2010-02-11 2013-03-19 Disney Enterprises, Inc. System and method for hybrid hierarchical segmentation
US9002859B1 (en) 2010-12-17 2015-04-07 Moonshadow Mobile, Inc. Systems and methods for high-speed searching and filtering of large datasets
WO2012097009A2 (en) * 2011-01-10 2012-07-19 Ward Roy W Systems and methods for high-speed searching and filtering of large datasets
EP2671168A4 (en) * 2011-02-03 2017-03-08 Voxeleron LLC Method and system for image analysis and interpretation
US9026591B2 (en) 2011-02-28 2015-05-05 Avaya Inc. System and method for advanced communication thread analysis
US9547693B1 (en) 2011-06-23 2017-01-17 Palantir Technologies Inc. Periodic database search manager for multiple data sources
US9026560B2 (en) * 2011-09-16 2015-05-05 Cisco Technology, Inc. Data center capability summarization
US8886651B1 (en) 2011-12-22 2014-11-11 Reputation.Com, Inc. Thematic clustering
US9171054B1 (en) 2012-01-04 2015-10-27 Moonshadow Mobile, Inc. Systems and methods for high-speed searching and filtering of large datasets
US8990204B1 (en) 2012-01-17 2015-03-24 Roy W. Ward Processing and storage of spatial data
US8751963B1 (en) 2013-01-23 2014-06-10 Splunk Inc. Real time indication of previously extracted data fields for regular expressions
US20140208217A1 (en) 2013-01-22 2014-07-24 Splunk Inc. Interface for managing splittable timestamps across event records
US8682906B1 (en) 2013-01-23 2014-03-25 Splunk Inc. Real time display of data field values based on manual editing of regular expressions
US9753909B2 (en) 2012-09-07 2017-09-05 Splunk, Inc. Advanced field extractor with multiple positive examples
US10394946B2 (en) 2012-09-07 2019-08-27 Splunk Inc. Refining extraction rules based on selected text within events
US8751499B1 (en) 2013-01-22 2014-06-10 Splunk Inc. Variable representative sampling under resource constraints
US9594814B2 (en) 2012-09-07 2017-03-14 Splunk Inc. Advanced field extractor with modification of an extracted field
US10013477B2 (en) 2012-11-19 2018-07-03 The Penn State Research Foundation Accelerated discrete distribution clustering under wasserstein distance
US9720998B2 (en) * 2012-11-19 2017-08-01 The Penn State Research Foundation Massive clustering of discrete distributions
US8977589B2 (en) 2012-12-19 2015-03-10 International Business Machines Corporation On the fly data binning
US8909642B2 (en) 2013-01-23 2014-12-09 Splunk Inc. Automatic generation of a field-extraction rule based on selections in a sample event
US9152929B2 (en) 2013-01-23 2015-10-06 Splunk Inc. Real time display of statistics and values for selected regular expressions
KR102029055B1 (en) * 2013-02-08 2019-10-07 삼성전자주식회사 Method and apparatus for high-dimensional data visualization
US10275778B1 (en) 2013-03-15 2019-04-30 Palantir Technologies Inc. Systems and user interfaces for dynamic and interactive investigation based on automatic malfeasance clustering of related data in various data structures
US8818892B1 (en) 2013-03-15 2014-08-26 Palantir Technologies, Inc. Prioritizing data clusters with customizable scoring strategies
US9965937B2 (en) 2013-03-15 2018-05-08 Palantir Technologies Inc. External malware data item clustering and analysis
CN103309940B (en) * 2013-05-03 2017-03-08 上海证券交易所 A kind of method to the sequence of out of order data flow
US9116975B2 (en) 2013-10-18 2015-08-25 Palantir Technologies Inc. Systems and user interfaces for dynamic and interactive simultaneous querying of multiple data stores
WO2015088780A1 (en) * 2013-12-10 2015-06-18 University Of Southern California Noise-enhanced clustering and competitive learning
US10579647B1 (en) 2013-12-16 2020-03-03 Palantir Technologies Inc. Methods and systems for analyzing entity performance
US10356032B2 (en) 2013-12-26 2019-07-16 Palantir Technologies Inc. System and method for detecting confidential information emails
US8832832B1 (en) 2014-01-03 2014-09-09 Palantir Technologies Inc. IP reputation
US9619557B2 (en) 2014-06-30 2017-04-11 Palantir Technologies, Inc. Systems and methods for key phrase characterization of documents
US9535974B1 (en) 2014-06-30 2017-01-03 Palantir Technologies Inc. Systems and methods for identifying key phrase clusters within documents
US9202249B1 (en) 2014-07-03 2015-12-01 Palantir Technologies Inc. Data item clustering and analysis
US9256664B2 (en) 2014-07-03 2016-02-09 Palantir Technologies Inc. System and method for news events detection and visualization
CN104156418B (en) * 2014-08-01 2015-09-30 北京系统工程研究所 The evolution clustering method that a kind of knowledge based is reused
US9501851B2 (en) 2014-10-03 2016-11-22 Palantir Technologies Inc. Time-series analysis system
US9767172B2 (en) 2014-10-03 2017-09-19 Palantir Technologies Inc. Data aggregation and analysis system
US10255345B2 (en) * 2014-10-09 2019-04-09 Business Objects Software Ltd. Multivariate insight discovery approach
US9984133B2 (en) 2014-10-16 2018-05-29 Palantir Technologies Inc. Schematic and database linking system
US9043894B1 (en) 2014-11-06 2015-05-26 Palantir Technologies Inc. Malicious software detection in a computing system
US10552994B2 (en) 2014-12-22 2020-02-04 Palantir Technologies Inc. Systems and interactive user interfaces for dynamic retrieval, analysis, and triage of data items
US9348920B1 (en) 2014-12-22 2016-05-24 Palantir Technologies Inc. Concept indexing among database of documents using machine learning techniques
US10362133B1 (en) 2014-12-22 2019-07-23 Palantir Technologies Inc. Communication data processing architecture
US9367872B1 (en) 2014-12-22 2016-06-14 Palantir Technologies Inc. Systems and user interfaces for dynamic and interactive investigation of bad actor behavior based on automatic clustering of related data in various data structures
US20160253672A1 (en) * 2014-12-23 2016-09-01 Palantir Technologies, Inc. System and methods for detecting fraudulent transactions
US9817563B1 (en) 2014-12-29 2017-11-14 Palantir Technologies Inc. System and method of generating data points from one or more data stores of data items for chart creation and manipulation
US10103953B1 (en) 2015-05-12 2018-10-16 Palantir Technologies Inc. Methods and systems for analyzing entity performance
US10191966B2 (en) * 2015-07-08 2019-01-29 Business Objects Software Ltd. Enabling advanced analytics with large data sets
US9454785B1 (en) 2015-07-30 2016-09-27 Palantir Technologies Inc. Systems and user interfaces for holistic, data-driven investigation of bad actor behavior based on clustering and scoring of related data
US9456000B1 (en) 2015-08-06 2016-09-27 Palantir Technologies Inc. Systems, methods, user interfaces, and computer-readable media for investigating potential malicious communications
US10489391B1 (en) 2015-08-17 2019-11-26 Palantir Technologies Inc. Systems and methods for grouping and enriching data items accessed from one or more databases for presentation in a user interface
US10102280B2 (en) * 2015-08-31 2018-10-16 International Business Machines Corporation Determination of expertness level for a target keyword
US10467204B2 (en) * 2016-02-18 2019-11-05 International Business Machines Corporation Data sampling in a storage system
US10521411B2 (en) 2016-08-10 2019-12-31 Moonshadow Mobile, Inc. Systems, methods, and data structures for high-speed searching or filtering of large datasets
US10318630B1 (en) 2016-11-21 2019-06-11 Palantir Technologies Inc. Analysis of large bodies of textual data
US10620618B2 (en) 2016-12-20 2020-04-14 Palantir Technologies Inc. Systems and methods for determining relationships between defects
US10691700B1 (en) * 2016-12-30 2020-06-23 Uber Technologies, Inc. Table replica allocation in a replicated storage system
US10325224B1 (en) 2017-03-23 2019-06-18 Palantir Technologies Inc. Systems and methods for selecting machine learning training data
US10606866B1 (en) 2017-03-30 2020-03-31 Palantir Technologies Inc. Framework for exposing network activities
US10579663B2 (en) * 2017-05-02 2020-03-03 International Business Machines Corporation Data insight discovery using a clustering technique
US10235461B2 (en) 2017-05-02 2019-03-19 Palantir Technologies Inc. Automated assistance for generating relevant and valuable search results for an entity of interest
US10482382B2 (en) 2017-05-09 2019-11-19 Palantir Technologies Inc. Systems and methods for reducing manufacturing failure rates
US10503526B2 (en) * 2017-06-13 2019-12-10 Western Digital Technologies, Inc. Method and system for user experience event processing and analysis
CN109685092B (en) * 2018-08-21 2024-02-06 中国平安人寿保险股份有限公司 Clustering method, equipment, storage medium and device based on big data
EP3623965A1 (en) 2018-09-14 2020-03-18 United Technologies Corporation Fabricated data detection method
CN110222030B (en) * 2019-05-13 2021-08-06 福建天泉教育科技有限公司 Dynamic database capacity expansion method and storage medium
US11551024B1 (en) * 2019-11-22 2023-01-10 Mastercard International Incorporated Hybrid clustered prediction computer modeling

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5706503A (en) * 1994-05-18 1998-01-06 Etak Inc Method of clustering multi-dimensional related data in a computer database by combining the two verticles of a graph connected by an edge having the highest score
US5832182A (en) * 1996-04-24 1998-11-03 Wisconsin Alumni Research Foundation Method and system for data clustering for very large databases
US5884305A (en) * 1997-06-13 1999-03-16 International Business Machines Corporation System and method for data mining from relational data by sieving through iterated relational reinforcement

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5706503A (en) * 1994-05-18 1998-01-06 Etak Inc Method of clustering multi-dimensional related data in a computer database by combining the two verticles of a graph connected by an edge having the highest score
US5832182A (en) * 1996-04-24 1998-11-03 Wisconsin Alumni Research Foundation Method and system for data clustering for very large databases
US5884305A (en) * 1997-06-13 1999-03-16 International Business Machines Corporation System and method for data mining from relational data by sieving through iterated relational reinforcement

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008034213A1 (en) * 2006-09-18 2008-03-27 Infobright Inc. A method and system for data compression in a relational database
US8266147B2 (en) 2006-09-18 2012-09-11 Infobright, Inc. Methods and systems for database organization
US8700579B2 (en) 2006-09-18 2014-04-15 Infobright Inc. Method and system for data compression in a relational database
US8838593B2 (en) 2006-09-18 2014-09-16 Infobright Inc. Method and system for storing, organizing and processing data in a relational database
WO2011056099A1 (en) 2009-11-03 2011-05-12 Telefonaktiebolaget L M Ericsson (Publ) Reducing computational complexity during user data analysis
EP2497256A1 (en) * 2009-11-03 2012-09-12 Telefonaktiebolaget LM Ericsson (publ) Reducing computational complexity during user data analysis
EP2497256A4 (en) * 2009-11-03 2012-10-31 Ericsson Telefon Ab L M Reducing computational complexity during user data analysis
US8417727B2 (en) 2010-06-14 2013-04-09 Infobright Inc. System and method for storing data in a relational database
US8521748B2 (en) 2010-06-14 2013-08-27 Infobright Inc. System and method for managing metadata in a relational database
US8943100B2 (en) 2010-06-14 2015-01-27 Infobright Inc. System and method for storing data in a relational database

Also Published As

Publication number Publication date
EP1062590A4 (en) 2004-05-12
DE69938339D1 (en) 2008-04-24
EP1062590B1 (en) 2008-03-12
US6374251B1 (en) 2002-04-16
DE69938339T2 (en) 2009-03-12
EP1062590A1 (en) 2000-12-27
ATE389213T1 (en) 2008-03-15

Similar Documents

Publication Publication Date Title
US6374251B1 (en) Scalable system for clustering of large databases
US6012058A (en) Scalable system for K-means clustering of large databases
US6263337B1 (en) Scalable system for expectation maximization clustering of large databases
US6449612B1 (en) Varying cluster number in a scalable clustering system for use with large databases
US6581058B1 (en) Scalable system for clustering of large databases having mixed data attributes
US6633882B1 (en) Multi-dimensional database record compression utilizing optimized cluster models
US7246125B2 (en) Clustering of databases having mixed data attributes
US6115708A (en) Method for refining the initial conditions for clustering with applications to small and large database clustering
US6263334B1 (en) Density-based indexing method for efficient execution of high dimensional nearest-neighbor queries on large databases
US6212526B1 (en) Method for apparatus for efficient mining of classification models from databases
Matignon Data mining using SAS enterprise miner
US6549907B1 (en) Multi-dimensional database and data cube compression for aggregate query support on numeric dimensions
CN111324642A (en) Model algorithm type selection and evaluation method for power grid big data analysis
US7873669B2 (en) Direct write back systems and methodologies
US8135667B2 (en) System, method, and computer-readable medium that facilitate in-database analytics with supervised data discretization
US7120624B2 (en) Optimization based method for estimating the results of aggregate queries
Aslam et al. Static and dynamic information organization with star clusters
CN109886334A (en) A kind of shared nearest neighbor density peak clustering method of secret protection
US20030229604A1 (en) Large margin perceptrons for document categorization
Hilderman et al. Measuring the interestingness of discovered knowledge: A principled approach
WO1992017853A2 (en) Direct data base analysis, forecasting and diagnosis method
EP1090362A1 (en) A scalable system for clustering of large databases having mixed data attributes
Mousavi Enhance online-offline varying density method for data stream clustering
Cha A context-aware similarity search for a handwritten digit image database
Kompalli Knowledge Discovery Using Data Stream Mining: An Analytical Approach

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): JP

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
WWE Wipo information: entry into national phase

Ref document number: 1999913907

Country of ref document: EP

WWP Wipo information: published in national office

Ref document number: 1999913907

Country of ref document: EP