On the Comparison of Semi-Supervised Hierarchical Clustering Algorithms in Text Mining Tasks

Semi-supervised clustering approaches have emerged as an option for enhancing clustering results. These algorithms use external information to guide the clustering process. In particular, semi-supervised hierarchical clustering approaches have been explored in many fields in the last years. These algorithms provide efficient and personalized hierarchical overviews of datasets. To the best of the authors’ knowledge, however, semi-supervised hierarchical clustering algorithms have not been extensively tested in document clustering. In this paper, we compare the performance in clustering document collections of three state of the art semi-supervised hierarchical clustering algorithms. These algorithms employ different ways of incorporating the external information in the clustering process: the CCL, which uses clusterlevel constraints based in Complete-Link approach; the HCAC, which uses cluster-level constraints based in the AverageLink approach; and the Pairwise-Constrained, which uses instance-level constraints based in the Average-Link approach. Experimental results indicate that using cluster-level or instance-level constraints achieve significant improvement in clustering quality when compared to unsupervised approaches. In particular, HCAC outperforms other methods with statistical significance.