Improving clustering with constrained communities

Abstract

In this paper, we propose a new constrained clustering algorithm, named Constrained Community Clustering (C3). It can utilize both must-link and cannot-link constraints that specify the pairs of data that belong to the same or different clusters. Instead of directly enforcing the pairwise constraints on the constrained data, C3 first builds constrained communities around each constrained data, and then exert pairwise constraints on the constrained communities. Therefore, C3 can not only extend the influence of pairwise constraints to the surrounding unconstrained data, but also uncover the underlying sub-structures of the clusters. The promising experimental results on the real-world text documents, handwritten digits, alphabetic characters, face recognition, and community discovery illustrate the effectiveness of our method.