A comparative study of semi-supervised clustering methods with pairwise constraint

Abstract

Abstract—Semi-Supervised Clustering (SSC) is a largely unsu- pervised learning task that seeks to guide the clustering process through constraint, and combines several methods with different approaches. In this work, our interest is more focused on the semi-supervised clustering with constraint approaches, and more particularly those based on the pairwise constraint. This paper establishes a comparative study between 3 algorithms mainly: the Constrained K-means algorithm which applies constraint of comparison between pairs of objects, called COP-KMEANS, the Semi-supervised kernel clustering with relative distance Algorithm (SKLR) and the Semi-supervised Kernel Mean Shift clustering Algorithm (SKMS). Experimental results indicate that the semi-supervised kernel Mean Shift clustering method can generally outperform the other semi-supervised methods. The experimental study shows that the use of constraint can improve performance especially when the number of available labeled examples is insufficient to build a decision model.