DualPop: User Popularity Preference Aware Debiasing Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommender systems usually suffer from popularity biases, where a few popular items are over-recommended, while most long-tail items struggle to gain exposure. Prior research focused on eliminating the influence of popularity to suppress popular items and enhance the recommendation of long-tail items. However, these debiasing methods face two limitations: (1) most strategies dogmatically eliminate the influence of popularity, without considering the differences in users’ tastes. This undermines the beneficial effect of popularity; and (2) they only consider static characteristics when modeling popularity, while ignoring the underlying temporal changes in popularity. Intuitively, popularity has a duality in recommendation and is not completely detrimental. Some users tend to follow trends, while others might have more niche tastes. Therefore, popularity is, to some extent, part of user preferences and helps achieve personalized recommendation. To overcome the aforementioned limitations and account for the duality in popularity, we propose a user popularity preference aware debiasing model, namely DualPop. It decouples the tangled representations of interest, popularity and detrimental popularity bias. Meanwhile, the representation of popularity is calibrated through exploiting its temporal features. Finally, we leverage counterfactual inference to achieve debiased recommendations. Extensive experiments on three real-world datasets show that DualPop outperforms six state-of-the-art baselines, demonstrating the effectiveness of considering the duality of popularity in recommendation.