K-medoid-style Clustering Algorithms for Supervised Summary Generation

This paper centers on the discussion of k-medoid-style clustering algorithms for supervised summary generation. This task requires clustering techniques that identify class-uniform clusters. This paper investigates such a novel clustering technique we term supervised clustering. Our work focuses on the generalization of k-medoid-style clustering algorithms. We investigate two supervised clustering algorithms: SRIDHCR (Single Representative Insertion/Deletion Hill Climbing with Restart) and SPAM, a variation of PAM. The solution quality and run time of these two algorithms as well as the traditional clustering algorithm PAM are evaluated using a benchmark consisting of four data sets. Experiments show that supervised clustering algorithms enhance class purity by 7% to 19% over the traditional clustering algorithm PAM, and that SRIDHCR finds better solutions than SPAM.