Research On Multi-View Discriminant Clustering Algorithms

Muti-view learning is the one of the hotspots in machine learning. It can be widely used in scene analysis, image processing, and web page information processing and other related fields. Based on single view discriminant clustering methods, we focus on multi-view clustering, and propose a series of multi-view clustering methods in this thesis. The main contributions of this thesis are listed as follows:(1) Firstly, we propose a multi-view discriminant clustering algorithm called MVDC-2 by applying dimension redunction and clustering methods in two view dataset, and then we extend it for n view cases. In adition, we propose a margin based multi-view discriminant clustering algorithm (MVDC-n). The expriment results validate the effectiveness of the proposed algorithms.(2) Secondly, kernel method is introduced into multi-view discriminant clustering algorithms. A kernelized multi-view discriminant clustering algorithm called KMDC and a kernelized margin based multi-view discriminant clustering algorithm called KMMDC are proposed separately. We validate the effectiveness of our algorithms in Handwriting dataset, ORL and WebKB dataset.(3) Thirdly, we introduce canonical correlation analysis (CCA) and its'discriminant format (including discriminant CCA and local discriminant CCA) into the framework of multi-view clustering algorithms, and three multi-view discriminant clustering algorithms are proposed. And we validate the effectiveness of those algorithms in actual datasets.