Self-Paced Multi-Task Multi-View Capped-norm Clustering

Yazhou Ren, Xin Yan, Zechuan Hu, Zenglin Xu

Abstract: Recently, multi-task multi-view clustering (MTMVC) which is able to utilize the relation of different tasks and the information from multiple views under each task to improve the clustering performance has attracted more and more attentions. However, MTMVC typically solves a non-convex optimization problem and thus is easy to stuck into bad local optima. In addition, noises and outliers generally have negative effects on the clustering performance. To alleviate these problems, we propose a novel self-paced multi-task multi-view capped-norm clustering (SPMTMVCaC) method, which progressively selects data samples to train the MTMVC model from simplicity to complexity. A novel capped-norm term is embedded into the objective of SPMTMVCaC model to reduce the negative influence of noises and outliers, and to further enhance the clustering performance. An efficient alternating optimization method is developed to solve the proposed model. Experimental results on real data sets demonstrate the effectiveness and robustness of the proposed method.