An adaptive semi-supervised clustering approach via multiple density-based information

Abstract

Since multimedia information has been dramatically increasing, multimedia data mining has drawn much more attentions than ever. As one of important mining tasks, clustering provided underpinning techniques for discovering the intrinsic structure and condensing information over large amount of multimedia data. Although many approaches have been proposed to improve performance and accuracy of clustering process in semi-supervised learning way, they are still quite sensitive to the assumptions of cluster structures and algorithm parameters setups. In this paper, we propose an adaptive semi-supervised clustering approach via multiple density-based information. It can automatically determine sets of important density-based parameters in use of both labeled and unlabeled data. Based on the different sets of density-based parameters, our approach is able to adaptively identify complex cluster structures in different size, shape and density without knowing the number of clusters; moreover, it is quite insensitive to the noise. Our approach has been generally evaluated on synthetic data and a collection of benchmark data sets, and yields promising clustering results. Furthermore, it has been also applied on Yale face database B, and simulation results specifically show its potential for practical application in face recognitions.