Prioritized Semi-supervised Deep Embedded Clustering

Abstract

Clustering, to group similar objects, is an important problem. Recently deep learning-based methods like Deep Embedded Clustering (DEC) [6] and its semi-supervised version called Semi-supervised Deep Embedded Clustering (SDEC) [12], where partially labeled data or data with constraints is available, are shown to give promising results. Both DEC and SDEC learn a latent space where similar objects are closer and dissimilar are away. While promising results are shown, the information present in constraints or a labeled subset of the data is not fully utilized. This paper proposes to use priorities for constraints so that important constraints are given more weightage than unimportant ones. Those constraints with points that are far away, but should be clustered into a group, gets more weight than other labeled points. Similarly, those in different groups which are very close get more weightage. The appropriate loss function is used in the learning process. The proposed method is called Prioritized Semi-supervised Deep Embedded Clustering (PSDEC). The results are compared using a few standard data sets against recent and classical similar methods. PSDEC is found to achieve a better result than un-prioritized constraints.