Adaptive Incremental Fusion With Global Semantics Alignment for Multi-View Representation Learning

Multi-view representation learning aims to exploit complementary and consensus information from heterogeneous data sources. However, existing approaches often suffer from unstable optimization and insufficient modeling of cross-view semantic dependencies. To this end, we propose AIGA, a novel framework that unifies global consistency and adaptive complementarity. AIGA first employs residual-based attention fusion to learn cross-view incremental representations while preserving the intrinsic characteristics of current view, enabling informative and discriminative local feature enhancement. Subsequently, G-Net is introduced to dynamically regulate the intensity of complementary information flow among attention-weighted views. By adaptively filtering out redundant content and emphasizing key complementary information, G-Net effectively balances the contribution of multiple views and facilitates consistent multi-view feature integration. In addition, theoretical analysis verifies the stability of the proposed fusion process. Optimized in an end-to-end manner under reconstruction, mutual information maximization, and contrastive learning objectives, AIGA produces stable, semantically consistent, and highly discriminative multi-view representations, achieving competitive performance compared with state-of-the-art methods across multiple benchmark datasets.