Semantic consistency enhancement and contribution-driven network for partial multi-view incomplete multi-label classification

Abstract

Abstract In recent years, multi-view multi-label learning has garnered considerable attention due to its broad application prospects, such as bioinformatics and medical imaging. However, the integrity of multi-view and multi-label data cannot be guaranteed in practical scenarios. Currently, some frameworks have been introduced to address the complex issue of partial multi-view incomplete multi-label classification, but they frequently overlook the impact of view quality on the learning of semantic information. To tackle this problem, we propose a semantic consistency enhancement and contribution-driven network(SCECD-Net). Different from the existing works, we focus on the quality differences between views, dynamically adjusting the mutual information learning process by quantifying view quality to reduce the model’s reliance on noisy information and more effectively capture view consistency. Furthermore, considering that treating each view equally during the reconstruction process may limit the model’s ability to leverage useful information, we propose a contribution-driven reconstruction strategy, which captures the contributions of each view using reliable supervisory information and employs this to balance the reconstruction of views, selectively retaining the most critical information. Extensive evaluations conducted on five datasets indicate that our method performs better than other approaches.

Keywords

• Partial multi-view
• Incomplete multi-label classification
• Mutual information learning
• Contribution-driven reconstruction