Go to ICLR 2026 Conference homepageLeveraging Holistic Explanations to Mitigate Popularity Bias for Recommender Systems
Keywords: Large Language Models, Recommender Systems, Explainable Recommendation, Debiased Recommender Systems
Abstract: Recommender systems often suffer from popularity bias, where items with high historical engagement ensure a dominant presence in the recommendation lists while equally relevant but less popular items (called niche items) remain under exposed towards majority of the users, thus impacting their reach within mainstream platforms. This bias arises partly due to the learning strategy of existing recommender models which display heavy reliance on interaction frequency and shallow contextual features that characterize any item, which fail to capture the true preferences of any users. To address this, we propose Expl-Debias, a novel framework that leverages holistic explanations to enrich user–item preference modeling and mitigate popularity bias. Expl-Debias operates in two stages: (Stage-1) a base training phase that learns general user–item utility, and (Stage-2) a contrastive explanation-aware training phase that incorporates LLM-generated positive and negative explanations to explicitly guide relevance learning toward personally aligned items and away from popular yet irrelevant ones. Extensive experiments on three real-world datasets demonstrate that our approach significantly improves recommendation accuracy while substantially reducing popularity bias, outperforming state-of-the-art LLM recommendation and debiasing baselines. These results demonstrate that integrating contrastive explanations offers an effective new direction for mitigating popularity bias in recommendation by balancing the tradeoff occurring between the recommendation performance and the negative effect of popularity bias. We provide our code at https://anonymous.4open.science/r/Expl-Pop-Bias-089A/.
Primary Area: other topics in machine learning (i.e., none of the above)
Submission Number: 16432
Loading