Techniques for improving clustering and association rules mining from very large transactional databases

Abstract

Clustering and association rules mining are two core data mining tasks that have been actively studied by data mining community for nearly two decades. Though many clustering and association rules mining algorithms have been developed, no algorithm is better than others on all aspects, such as accuracy, efficiency, scalability, adaptability and memory usage. While more efficient and effective algorithms need to be developed for handling the large-scale and complex stored datasets, emerging applications where data takes the form of streams pose new challenges for the data mining community. The existing techniques and algorithms for static stored databases cannot be applied to the data streams directly. They need to be extended or modified, or new methods need to be developed to process the data streams.In this thesis, algorithms have been developed for improving efficiency and accuracy of clustering and association rules mining on very large, high dimensional, high cardinality, sparse transactional databases and data streams.A new similarity measure suitable for clustering transactional data is defined and an incremental clustering algorithm, INCLUS, is proposed using this similarity measure. The algorithm only scans the database once and produces clusters based on the user’s expectations of similarities between transactions in a cluster, which is controlled by the user input parameters, a similarity threshold and a support threshold. Intensive testing has been performed to evaluate the effectiveness, efficiency, scalability and order insensitiveness of the algorithm.To extend INCLUS for transactional data streams, an equal-width time window model and an elastic time window model are proposed that allow mining of clustering changes in evolving data streams. The minimal width of the window is determined by the minimum clustering granularity for a particular application. Two algorithms, CluStream_EQ and CluStream_EL, based on the equal-width window model and the elastic window model respectively, are developed by incorporating these models into INCLUS. Each algorithm consists of an online micro-clustering component and an offline macro-clustering component. The online component writes summary statistics of a data stream to the disk, and the offline components uses those summaries and other user input to discover changes in a data stream. The effectiveness and scalability of the algorithms are evaluated by experiments.This thesis also looks into sampling techniques that can improve efficiency of mining association rules in a very large transactional database. The sample size is derived based on the binomial distribution and central limit theorem. The sample size used is smaller than that based on Chernoff Bounds, but still provides the same approximation guarantees. The accuracy of the proposed sampling approach is theoretically analyzed and its effectiveness is experimentally evaluated on both dense and sparse datasets.Applications of stratified sampling for association rules mining is also explored in this thesis. The database is first partitioned into strata based on the length of transactions, and simple random sampling is then performed on each stratum. The total sample size is determined by a formula derived in this thesis and the sample size for each stratum is proportionate to the size of the stratum. The accuracy of transaction size based stratified sampling is experimentally compared with that of random sampling.The thesis concludes with a summary of significant contributions and some pointers for further work.