Partial Multi-View Incomplete Multi-Label Learning Network With Quality-Aware Representation Fusion

Recently, the topic of multi-view multi-label classification has aroused significant attention from scholars. Plenty of methods adopt an average weighting scheme to merge the features obtained from multiple views, which commonly ignore the quality difference of information provided by multiple views and thus limit the credibility of the fusion feature for the overall task. Besides, most of these methods assume the views and labels are complete while neglecting both views and labels may be incomplete. To solve these problems, we propose a quality-aware representation fusion network for partial multi-view incomplete multi-label classification, named QARF-net. Since assigning equal fusion weights for each view may be not in line with the actual contributions of individual views, a view quality-aware module is proposed to learn suitable weights for different views dynamically based on the quality of each view’s information, which provides a reliable guide for fusing the information of multiple views. In addition, considering the consistency characteristics of multi-view data, we impose a sample-level dual constraint to preserve the consistency property of the feature in multi-view space and constrain the sample structure in the fused feature space, respectively. Last but not least, QARF-net can not only deal with complete multi-view multi-label classification tasks but also tackle partial multi-view incomplete multi-label classification tasks. Experimental results on five real-world datasets indicate that our proposed method outperforms state-of-the-art methods.