Clustering with Extended Constraints by Co-Training

Constrained Clustering is a data mining technique that produces clusters of similar data by using pre-given constraints about data pairs. If we consider using constrained clustering for some practical interactive systems such as information retrieval or recommendation systems, the cost of constraint preparation will be the problem as well as other machine learning techniques. In this paper, we propose a method to complement the lack of constraints by using co-training framework, which extends training examples by leveraging two kinds of feature sets. Our method is based on a constrained clustering ensemble algorithm that integrates a set of clusters produced by a constrained k-means with random ordered data assignment, and runs the same algorithm on two different feature sets to extend constraints. We evaluate our method on a Web page dataset that provides two different feature sets. The results show that our method achieves the performance improvement by using co-training approach.