Semi-supervised clustering using incomplete prior knowledge

Abstract

Clustering algorithms incorporated with prior knowledge have been widely studied and many nice results were shown in recent years. However, most existing algorithms implicitly assume that the prior information is complete, typically specified in the form of labeled objects with each category. These methods decay and behave unstably when the labeled classes are incomplete. In this paper a new type of prior knowledge which bases on partially labeled data is proposed. Then we develop two novel semi-supervised clustering algorithms to face this new challenge. An empirical study performed on benchmark dataset shows that our proposed algorithms produce better results with limited labeled examples comparing with existing baselines. © Springer-Verlag Berlin Heidelberg 2007.