Discrepancy-Aware Contrastive Multi-View Representation Fusion With Truncated Cross Attention for Mitigating Performance Degeneration

Multi-view representation fusion aims to enhance fusion representation performance by integrating information from multiple views. Since diverse views may have significant quality discrepancies, traditional representation fusion methods often neglect the distribution gap between high-quality and low-quality views. Recent advances have seen the rise of attention mechanisms in multi-view representation fusion, but these methods struggle with inter-view interactions and noise distillation. To this end, we propose the Discrepancy-aware Contrastive Multi-view Representation Fusion with Truncated Cross Attention (DCMRF-TCA) network. Our approach leverages an entropy-guided truncated cross attention mechanism to selectively fuse views with minimal discrepancies, enhancing the discriminability of the fusion representation. Additionally, we introduce a discrepancy-aware auto-weighted contrastive learning objective that adjusts contrastive weights based on total view discrepancies, mitigating the negative impact of low-quality views. Extensive experiments on public multi-view datasets validate the effectiveness of our framework in handling view discrepancies and mitigating performance degeneration in multi-view representation fusion.