Clustering in applications with multiple data sources-A mutual subspace clustering approach

摘要

In many applications, such as bioinformatics and cross-market customer relationship management, there are data from multiple sources jointly describing the same set of objects. An important data mining task is to find interesting groups of objects that form clusters in subspaces of the data sources jointly supported by those data sources. In this paper, we study a novel problem of mining mutual subspace clusters from multiple sources. We develop two interesting models and the corresponding methods for mutual subspace clustering. The density-based model identifies dense regions in subspaces as clusters. The bottom-up method searches for density-based mutual subspace clusters systematically from low-dimensional subspaces to high-dimensional ones. The partitioning model divides points in a data set into k exclusive clusters and a signature subspace is found for each cluster, where k is the number of clusters desired by a user. The top-down method interleaves the well-known k-means clustering procedures in multiple sources. We use experimental results on synthetic data sets and real data sets to report the effectiveness and the efficiency of the methods.