PAD: Popularity-aware debiasing for high-value item recommendation

Abstract

Recommender systems play a crucial role in our daily lives. However, in the context of high-value item recommendation, they face significant challenges. Due to the high price of these items, user purchase histories are often extremely sparse, making it difficult for recommender systems to accurately capture user preferences. Consequently, they tend to over-rely on popularity information. Moreover, the high-value item market exhibits a pronounced imbalanced distribution, where most user interactions focus on popular items. As a result, traditional recommender systems tend to prioritize these items while rarely recommending less popular ones, leading to low recommendation coverage. To address this challenge, we propose a Popularity-Aware Debiasing (PAD) model, which improves recommendation coverage in high-value item scenarios without compromising accuracy. First, we employ soft prompts to guide a pre-trained language model (PLM) in enriching user representations. By incorporating semantic knowledge from the PLM, our model captures more comprehensive user preferences, ensuring recommendation accuracy while mitigating the model’s dependence on popularity signals. Building upon this, we apply popularity-aware debiasing to reduce overfitting and enhance coverage. PAD prevents the recommendation model from indiscriminately recommending the most popular items to all users, encouraging it to explore a wider range of items in its recommendations. Experiments conducted on industrial and public datasets demonstrate that our method mitigates popularity bias, significantly improving item recommendation coverage while maintaining accuracy.