Collaborative Research: Data Mining: Theory and Algorithms

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

This collaborative research project is carried out by Leonard Pitt at the University of Illinois and H. V. Jagadish at the University of Michigan, drawing on Professor Pitt's experience in computational learning theory, and Professor Jagadish's expertise in databases. The goal of this research is to develop a theoretical framework for investigating data mining problems, and to adapt existing machine learning algorithms and develop new ones within the framework. Typically, the focus of Machine Learning algorithms is on learning a single classifier that works well for most of the (labeled) data, whereas data mining focuses on learning "heuristic" rules, typically on unlabeled data, that give insight into the nature of the data. By unifying techniques from the former area with goals of the latter, new clustering algorithms are developed that deal with massive data sets via sampling, yet provide optimality guarantees. The research also provides criteria that data mining practitioners may apply in deciding whether aggregation or sampling is a preferred data reduction technique for the task at hand. Algorithms for new and useful types of data patterns are designed, and algorithms that incorporate the user into the data-mining task are analyzed and developed within the framework.