A multi-view multi-label learning with incomplete data and self-adaptive correlations

Abstract Multi-view multi-label (MVML) learning aims to process MVML data sets represented with multiple feature sets (i.e., views) and labeled with multiple class labels. While in current scenes, MVML data sets always encounter to two main phenomena. First, for MVML data, there exist some correlations among different features, instances, labels and these correlations usually have diverse representations including within-view, cross-view, and consensus-view representations. Due to there are usually some certain relationships between information exist, thus these correlations always be changed in a self-adaptive way. Second, for some unpredictable reasons, MVML data maybe incomplete and loss some information. To address these phenomena, we pay attention to the self-adaptive measurement of those correlations in different representations and the process of incomplete data, then a multi-view multi-label learning with incomplete data and self-adaptive correlations (MVML-IDSaC) is developed. Extensive experiments on 5 MVML data sets show the superiority of the developed algorithm and some conclusions are addressed. (1) MVML-IDSaC performs better than some related competitive algorithms in statistical over AUC and precision; (2) MVML-IDSaC can process incomplete MVML data much better; (3) considering comprehensive relationships about data and its inferring results with a feasible way, the performances of a multi-view multi-label algorithm is promoted further.