A Comparison of Three Document Clustering Algorithms: TreeCluster, Word Intersection GQF, and Word Intersection Hierarchical Agglomerative Clustering

This work investigated three techniques to automatically cluster a collection of documents: Word-Intersection with GQF, Word-Intersection with hierarchical agglomerative clustering, and TreeClustering. The Word-Intersection algorithms have been previously described in the literature while the TreeClustering technique is novel to this work. The TreeCluster algorithm idea comes from rule induction techniques and is used to generate a shallow tree of clusters that a user can browse. This algorithm is also O(n) when used with a fixed tree depth, as opposed to O(n) as the other two algorithms. Experimental results on a collection of Mail and Web documents indicate that the agglomerative clustering algorithm performed the best, but also the slowest. The GQF algorithm performed well on the Mail domain, but performed poorly on the Web domain without tuning to its heuristic. The TreeCluster algorithm performed reasonably well on both domains, and was also the fastest algorithm out of the three algorithms tested.