Pairwise Constraint Propagation for Graph-Based Semi-supervised Clustering

Abstract

This paper proposes an approach for pairwise constraint propagation in graph-based semi-supervised clustering. In our approach, the entire dataset is represented as an edge-weighted graph by mapping each data instance as a vertex and connecting the instances by edges with their similarities. Based on the pairwise constraints, the graph is then modified by contraction in graph theory to reflect must-link constraints, and graph Laplacian in spectral graph theory to reflect cannot-link constraints. However, the latter was not effectively utilized in previous approaches. We propose to propagate pairwise constraints to other pairs of instances over the graph by defining a novel label matrix and utilizing it as a regularization term. The proposed approach is evaluated over several real world datasets, and compared with previous regularized spectral clustering and other methods. The results are encouraging and show that it is worthwhile to pursue the proposed approach.