On Finding Large Conjunctive Clusters

We propose a new formulation of the clustering problem that differs from previous work in several aspects. First, the goal is to explicitly output a collection of simple and meaningful conjunctive descriptions of the clusters. Second, the clusters might overlap, i.e., a point can belong to multiple clusters. Third, the clusters might not cover all points, i.e., not every point is clustered. Finally, we allow a point to be assigned to a conjunctive cluster description even if it does not completely satisfy all of the attributes, but rather only satisfies most. A convenient way to view our clustering problem is that of finding a collection of large bicliques in a bipartite graph. Identifying one largest conjunctive cluster is equivalent to finding a maximum edge biclique. Since this problem is NP-hard [28] and there is evidence that it is difficult to approximate [12], we solve a relaxed version where the objective is to find a large subgraph that is close to being a biclique. We gi...

• Type: Conference
• Year: 2003
• Where: COLT
• Authors: Nina Mishra, Dana Ron, Ram Swaminathan
• Keywords: Clustering Problem, COLT 2003, Conjunctive Cluster, Meaningful Conjunctive Descriptions