Web Document Clustering with Multi-view Information Bottleneck

Clustering is an important way to organize the large amount of information on the Web. In this paper, we study how to incorporate many information of Web document, such as content, anchor, URL etc, to improve the performance of clustering. We propose a novel algorithm: multi-view information bottleneck (MVIB), to cluster Web documents with multi-type features. In this algorithm, the compatible constraint maximizing the agreement between clustering hypotheses on different views is imposed on the individual views to cluster instances. Based on the compatible constraints, the set of clustering hypotheses revealing lots of information about correct one is obtained. The final hypothesis can be deduced from these hypotheses. We study the performance of MVIB in different views setting. Experiments on two real datasets indicate that MVIB with 3-view setting based on content, anchor text and URL can improve the quality of clusters more effectively.