Performance Evaluation of Constraints in Graph-Based Semi-supervised Clustering

Semi-supervised learning has been attracting much interest to cope with vast amount of data. When similarities among instances are specified, by connecting each pair of instances with an edge, the entire data can be represented as an edge-weighted graph. Based on the graph representation, we have proposed a graph-based approach for semi-supervised clustering, which modifies the graph structure by contraction in graph theory and graph Laplacian in spectral graph theory. In this paper we conduct extensive experiments over various document datasets and report its performance evaluation, with respect to the type of constraints as well as the number of constraints. We also compare it with other state of the art methods in terms of accuracy and running time, and the results are encouraging. Especially, our approach can leverage small amount of pairwise constraints to increase the performance.