Hierarchical Fusion with Dual Contrast for Incomplete Multi-View Representation Learning

Xiaojian Ding, Lin Zhao, Xiaoying Zhu, Xian Li

Keywords: Incomplete multi-view clustering;Dendritic fusion;Hierarchical dual contrast

TL;DR: DFHDC

Abstract:

Real-world multi-view data often suffers from partial incompleteness, posing challenges to traditional clustering methods. Existing approaches face two key limitations: (1) rigid fusion strategies that fail to dynamically select complementary views, and (2) contrastive learning methods that struggle to capture high-order dependencies and suppress intra-view redundancy. To address these issues, we propose DFHDC, a novel framework integrating dendritic fusion and hierarchical dual contrast mechanisms to dynamically select optimal
view combinations and construct multi-level semantic fusion pathways. The dendritic fusion strategy progressively fuses views in a bottom-up manner to maximize inter-view complementarity, while the hierarchical dual contrast mechanism performs contrastive learning in both local and global semantic spaces, simultaneously maximizing cross-view mutual information and minimizing intra-view redundancy, thereby enhancing the consistency and discriminability of the learned representations. Additionally, the framework incorporates a view-specific fine-tuning strategy to implicitly recover missing views. Experiments show DFHDC outperforms state-of the-art methods, especially under high missing rates, validating its effectiveness in incomplete multi-view learning.

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

Submission Number: 3237