Skip to content
Book Open access

Adaptive Preference Allocation and Semantic Hard-Negative Supervision for Multimodal Recommendation

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 27 references

Abstract

Multimodal recommendation enhances user preference modeling by leveraging visual and textual content. However, existing methods generally adopt a unified modality fusion strategy, making it difficult to capture users' diverse reliance on visual and textual signals. Meanwhile, random negative sampling provides limited ranking supervision and is insufficient for modeling fine-grained preference boundaries. To address these issues, we propose an Adaptive Preference allocation and Semantic Negative supervision framework called APSN, which explicitly models personalized preferences through user-level modality contribution allocation, thereby learning more fine-grained multimodal user representations. To enhance ranking supervision, APSN samples informative semantic negatives from dynamically constructed user communities, guiding the model to better capture users' preference boundaries. In addition, APSN improves feature robustness via masked semantic consistency and complements local structural information through a sparse collaborative view. Experimental results on multiple public datasets demonstrate that APSN can achieve consistent and effective improvements in recommendation performance. Code is available at https://github.com/jcs12123/APSN.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.