Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias

TL;DR

This work tackles popularity bias in collaborative filtering by introducing PAAC, a two-pronged approach that combines popularity-aware supervised alignment to bolster unpopular-item representations with a re-weighted contrastive learning scheme to reduce embedding separation caused by popularity. Built on a LightGCN backbone, PAAC jointly optimizes the standard CF loss with the proposed alignment and contrastive objectives, enabling more balanced representations without sacrificing overall accuracy. Extensive experiments on three real-world datasets with unbiased test splits show that PAAC achieves state-of-the-art performance, notably improving unpopular-item recommendations and narrowing the popularity gap, while providing insights through ablations and hyperparameter analyses. The method offers practical, tunable mechanisms (e.g., $\gamma$, $\beta$, and batch-ratio $x$) to adapt debiasing to different data distributions, with code available for reproduction.

Abstract

Collaborative Filtering (CF) typically suffers from the significant challenge of popularity bias due to the uneven distribution of items in real-world datasets. This bias leads to a significant accuracy gap between popular and unpopular items. It not only hinders accurate user preference understanding but also exacerbates the Matthew effect in recommendation systems. To alleviate popularity bias, existing efforts focus on emphasizing unpopular items or separating the correlation between item representations and their popularity. Despite the effectiveness, existing works still face two persistent challenges: (1) how to extract common supervision signals from popular items to improve the unpopular item representations, and (2) how to alleviate the representation separation caused by popularity bias. In this work, we conduct an empirical analysis of popularity bias and propose Popularity-Aware Alignment and Contrast (PAAC) to address two challenges. Specifically, we use the common supervisory signals modeled in popular item representations and propose a novel popularity-aware supervised alignment module to learn unpopular item representations. Additionally, we suggest re-weighting the contrastive learning loss to mitigate the representation separation from a popularity-centric perspective. Finally, we validate the effectiveness and rationale of PAAC in mitigating popularity bias through extensive experiments on three real-world datasets. Our code is available at https://github.com/miaomiao-cai2/KDD2024-PAAC.

Figures (6)

• Figure 1: Popularity bias presents two challenges: (1) Overfitting caused by limited supervisory signals for unpopular items, and (2) Representation separation in item embeddings driven by popularity bias.
• Figure 2: An Illustration of our proposed Popularity-Aware Alignment and Contrast (PAAC), which consists of the Supervised Alignment Module and the Re-weighting Contrast Module. Supervised Alignment Module leverages the common supervision signal in popular representations to guide the learning of unpopular representations. Re-weighting Contrast Module address representation separation from a popularity-centric perspective.
• Figure 3: Performance comparison over different item popularity groups. In particular, $\Delta$ indicates the accuracy gap between different groups.
• Figure 4: Performance comparison w.r.t. $\lambda_{1}$ and $\lambda_{2}$ on the Yelp2018 and Gowalla dataset in $NDCG@20$. The values indicate the percentage improvement relative to the best baseline.
• Figure 5: Performance comparison w.r.t. different $\gamma$ and $\beta$. The top shows the $NDCG@20$ and $HR@20$ results on Yelp2018 and the bottom shows the results on Gowalla. The horizontal line represents the best results already achieved in the baseline.