Machine Learning and Data Mining in Pattern Recognition

Abstract

In this paper we present a Pareto based multi objective algorithm for semi supervised clustering (PSC). Semi-supervised clustering uses a small amount of supervised data known as constraints, to assist unsupervised learning. Instead of modifying the clustering objective function, we add another objective function to satisfy specified constraints. We use a lexicographically ordered cluster assignment step to direct the search and a Pareto based multi objective evolutionary algorithm to maintain diversity in the population. Two objectives are considered: one that minimizes the intra cluster variance and another that minimizes the number of constraint violations. Experiments show the superiority of the method over a greedy algorithm (PCK-means) and a genetic algorithm (COP-HGA). © 2012 Springer-Verlag.