Recommendation systems are pivotal in personalizing user experiences across e-commerce platforms. Yet, they often suffer from inherent biases, such as popularity bias. Popularity bias refers to the tendency of recommendation systems to favor overrepresented items, leading to homogenized suggestions that limit diversity and fail to address the needs of users with unique or less common preferences. While large language models (LLMs) promise to enhance recommendation quality through semantic understanding, they struggle with biased data and insufficient collaborative knowledge. To address these challenges, we propose Causal-Enhanced LLM-based Debiased Recommender System (CLD-Rec), a novel two-stage framework that synergizes causal inference and LLMs to mitigate biases and improve recommendation robustness. In the first stage, causal reasoning identifies and eliminates biases (e.g., popularity bias) from user–item interactions, thereby generating debiased collaborative knowledge. In the second stage, this knowledge is integrated with LLMs; by leveraging the latter’s robust language comprehension capabilities, the system generates personalized recommendations that not only align with user preferences but also ensure fairness. Experimental results show that CLD-Rec consistently outperforms state-of-the-art models across multiple datasets, including Games, Toys, and Sports. Additionally, CLD-Rec demonstrates superior fairness by effectively mitigating popularity bias, leading to more balanced and diverse recommendations.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号