Multi-view multi-label-based online method with threefold correlations and dynamic updating multi-region

Changming Zhu, Shuaiping Guo, Dujuan Cao, YiTing Zhou, Duoqian Miao, Witold Pedrycz

Abstract: Semi-supervised real-time generation multi-view multi-label data sets are widely encountered in practical applications. A key issue is how to process the data whose information including labels or features may be lost due to some unforeknowable factors. In our work, we develop a multi-view multi-label-based online method with threefold correlations and dynamic updating multi-region (M2CR) to solve this issue. First, we adopt three kinds of correlations between features and labels to recover the missing information. Second, we process new arriving instances with dynamic updating multi-region. Experiments on classical multi-view multi-label data sets validate the effectiveness of M2CR in terms of classification, time performance, convergence, etc.