Self-Adaptive Within- and Consensus-View Correlation-Based Multi-View Multi-Label Learning

In current real-world applications, many algorithms are developed to process multi-view multi-label, multi-view, and multi-label data. However, they cannot express correlations among different instances, features, and labels in within-view and consensus-view representations self-adaptively and relatively accurately. To this end, this study takes the classical multiple correlations-based algorithm as the basis and explores some inherent laws of self-adaptive change for those correlations in within-view and consensus-view representations. The developed algorithm is named self-adaptive within- and consensus-view correlation-based multi-view multi-label learning (SWvCv-MVML). Compared with traditional algorithms, SWvCv-MVML has some highlights: 1) it utilizes the inherent laws of self-adaptive change for within-view and consensus-view correlation information to design a model and 2) it adopts matrix max-norm and detects the redundant elements for the feature matrices to avoid the over-fitting problems. Extensive experiments on 36 classical datasets validate the superiority of our algorithm in multiple aspects and some conclusions are addressed: 1) SWvCv-MVML outperforms most compared algorithms statistically in terms of AUC and its performance is also stable; 2) although the model of SWvCv-MVML is more complicated, its computational cost is still be moderate, and, on most datasets, SWvCv-MVML has a relatively fast convergence; and 3) introducing some laws of self-adaptive change for those correlations can improve the ability of SWvCv-MVML to process multi-view multi-label datasets effectively.