Selecting the Right Features for Bipartite-Based Text Clustering

Document datasets can be described with a bipartite graph where terms and documents are modeled as vertices on two sides respectively. Partitioning such a graph yields a co-clustering of words and documents, in the hope that the cluster topic can be captured by the top terms and documents in the same cluster. However, single terms alone are often not enough to capture the semantics of documents. To that end, in this paper, we propose to employ hyperclique patterns of terms as additional features for document representation. Then we use F-score to select the top discriminative features to construct the bipartite. Finally, the extensive experiments indicated that compared to the standard bipartite formulation, our approach is able to achieve better clustering performance at a smaller graph size.