Theory-Inspired Task-Relevant Representation Learning for Incomplete Multi-View Multi-Label Learning

Tianxiang Xu, Chang LIU, Yilun Wang, Li Xiyu

Keywords: Incomplete multi-view learning, Multi-label classification, Representation learning, Label semantics learning

Abstract:

Multi-view multi-label learning is commonly hindered by dual data incompleteness, arising from constraints in feature collection and prohibitive annotation costs. To address the intricate yet highly practical challenges and enhance the reliability of representation extraction, heterogeneous feature fusion, and label semantic learning, we propose a Theory-Inspired Task-Relevant Representation Learning method named TITRL.
From an information-theoretic standpoint, we identify the sources of view-specific information that interfere with shared representations. By introducing dual-layer constraints on feature exclusivity and label integration, TITRL constructs a general framework for task-relevant information extraction. Besides, through variational derivation, we demonstrate the existence of tractable bounds for the mutual information model that guides the optimization direction. Regarding label semantic learning, we establish flexible relationships between label prototypes by promoting the expression of sample-level label correlations.
During the multi-view integration process , TITRL simultaneously incorporates early fusion through distribution information aggregation and late fusion weighted by prediction confidence, which improves the semantic stability while enabling dynamic view quality assessment. Finally, extensive experimental results validate the effectiveness of TITRL against state-of-the-art methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 4652