PAMS-GNN: Popularity-Aware Multimodal Semantic GraphNeural Networks for Recommendation
作者
Qi Li,Y J Cao,Le Yu,Weiwei Li,Jin Xu
出处
期刊:International Conference on Multimedia Retrieval [Association for Computing Machinery] 日期:2026-06-15卷期号:: 59-68
标识
DOI:10.1145/3805622.3810751
摘要
Multimodal graph-based recommendation has shown strong potential by jointly modeling user–item interactions and rich multimodal item content. However, existing methods often suffer from noisy features, severe popularity bias, and uniform message propagation that fails to adapt to heterogeneous item popularity and varying signal reliability. To address these issues, we propose PAMS-GNN, a Popularity-Aware Multimodal Semantic Graph Neural Network that formulates recommendation as conditional message passing over collaborative and semantic graphs. The core idea is to adapt both propagation behaviors and channel reliability according to item popularity and local structural conditions. PAMS-GNN introduces a compact prototype-based framework with popularity-aware routing, enabling items from different popularity regimes to adopt distinct multi-hop propagation behaviors with minimal overhead. In addition, a variance-guided fusion mechanism adaptively balances collaborative and semantic signals based on local reliability, while lightweight projection and gating are used to purify multimodal representations and stabilize semantic graph construction. Extensive experiments on multiple benchmarks demonstrate that PAMS-GNN consistently outperforms state-of-the-art methods, effectively alleviating popularity bias and improving long-tail recommendation performance with favorable efficiency.