On semi-supervised clustering

Due to its capability to exploit training datasets encompassing both labeled and unlabeled patterns, semi–supervised learning (SSL) has been receiving attention from the community throughout the last decade. Several SSL approaches to data clustering have been proposed and investigated, as well. Unlike typical SSL setups, in semi–supervised clustering (SSC) the partial supervision is generally not available in terms of class labels associated with a subset of the training sample. In fact, general SSC algorithms rely rather on additional constraints which bring some kind of a–priori, weak side–knowledge to the clustering process. Significant instances are: COP–COBWEB and COP k–means, HMRF k–means, seeded k–means, constrained k–means, and active fuzzy constrained clustering. This chapter is a survey of major SSC philosophies, setups, and techniques. It provides the reader with an insight into these notions, categorizing and reviewing the major state–of–the–art approaches to SSC.