Co-clustering by similarity refinement

When a data set contains objects of multiple types, to cluster the objects of one type, it is often necessary to consider the cluster structure on the objects of the other types. Co-clustering the related objects often generates better clusters. One basic connection here is that the similarity among the objects of one type is often affected by the cluster structures on the objects of the other types. Although many co-clustering schemes have been proposed, none has explicitly explored such a connection. We propose a framework that utilizes this connection directly. In this framework, employing a spectral-embedding-based approach, we first obtain certain approximate cluster information about the objects of individual types. Such information is then used to refine the similarity measures for the objects of the related types. The final clustering is performed with the refined similarity. We tested our framework on both bipartite-graph and document clustering. Our experiments showed that the refined similarity leads to much better clustering, indicating that our refinement makes the similarity measures closer to their true values.