Scholarlink

<p>Popularity bias is a common problem in recommendation systems, where popular items are frequently over-recommended to users due to the cumulative advantages of exposure and interaction data. This bias primarily arises from two factors: first, popularity factors are highly entangled with interest factors, leading recommendation models to tend to overly rely on popularity signals, thereby ignoring users’ true interest needs. Second, current methodologies predominantly depend on user-item interaction data, which results in an inadequate utilization of the semantic information inherent in item content features. To tackle these challenges, we propose a new model named Disentangling Interest and Popularity Representation for Multimodal Recommendation, MR-DIP for short, to alleviate popularity bias in recommendation systems. First, we develop a popularity-aware module to evaluate the global popularity of items and the sensitivity of users to popular items, respectively. On this basis, MR-DIP constructs a dual-channel disentangling representation learning framework to separate intrinsic interests and popularity-driven factors implicit in the interaction relationships. To explore the influence of popularity factors, we analyse the content characteristics of items by extracting their visual and textual features. These features were aligned with the learned interests and popularity representations. Then, we establish a contrastive learning task to enhance the learned representations, encouraging interest factors and popularity factors to be disentangled from each other in the latent embedding space. Finally, a dynamic weighting aggregation mechanism is designed to retain the valid information in popularity signals while suppressing the popularity bias. Experimental studies demonstrate that MR-DIP significantly outperforms the state-of-the-art baseline models on various benchmark datasets. Ablation studies further validate its effectiveness in learning disentangled representations from interest and popularity perspectives.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Alleviating popularity Bias for multimodal recommendation from disentangling perspective

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Hangyuan Du,
  • Di Wu,
  • Liu Yang,
  • Weihong Zhang,
  • Yu Xie

摘要

Popularity bias is a common problem in recommendation systems, where popular items are frequently over-recommended to users due to the cumulative advantages of exposure and interaction data. This bias primarily arises from two factors: first, popularity factors are highly entangled with interest factors, leading recommendation models to tend to overly rely on popularity signals, thereby ignoring users’ true interest needs. Second, current methodologies predominantly depend on user-item interaction data, which results in an inadequate utilization of the semantic information inherent in item content features. To tackle these challenges, we propose a new model named Disentangling Interest and Popularity Representation for Multimodal Recommendation, MR-DIP for short, to alleviate popularity bias in recommendation systems. First, we develop a popularity-aware module to evaluate the global popularity of items and the sensitivity of users to popular items, respectively. On this basis, MR-DIP constructs a dual-channel disentangling representation learning framework to separate intrinsic interests and popularity-driven factors implicit in the interaction relationships. To explore the influence of popularity factors, we analyse the content characteristics of items by extracting their visual and textual features. These features were aligned with the learned interests and popularity representations. Then, we establish a contrastive learning task to enhance the learned representations, encouraging interest factors and popularity factors to be disentangled from each other in the latent embedding space. Finally, a dynamic weighting aggregation mechanism is designed to retain the valid information in popularity signals while suppressing the popularity bias. Experimental studies demonstrate that MR-DIP significantly outperforms the state-of-the-art baseline models on various benchmark datasets. Ablation studies further validate its effectiveness in learning disentangled representations from interest and popularity perspectives.