Mitigating Recommendation Biases via Group-Alignment and Global-Uniformity in Representation Learning

TL;DR

The paper tackles item-popularity and user-consistency biases in collaborative filtering by reframing debiasing as a representation-distribution problem. It introduces AURL, a CF-backbone-agnostic framework with two regularizers: Group-Alignment (via Maximum Mean Discrepancy) and Global-Uniformity (via Gaussian kernel on the unit sphere) to reduce group-discrepancy and global-collapse. Empirical results across Amazon-Book, Movielens-20M, and Douban-Book on BPRMF, LightGCN, and SimGCL show substantial bias reductions (PRU, DP) while maintaining competitive accuracy (HR@K, NDCG@K) and reveal the benefits of jointly optimizing both regularizers. These findings highlight the practical impact of shaping representation distributions for debiasing in real-world recommender systems.

Abstract

Collaborative Filtering~(CF) plays a crucial role in modern recommender systems, leveraging historical user-item interactions to provide personalized suggestions. However, CF-based methods often encounter biases due to imbalances in training data. This phenomenon makes CF-based methods tend to prioritize recommending popular items and performing unsatisfactorily on inactive users. Existing works address this issue by rebalancing training samples, reranking recommendation results, or making the modeling process robust to the bias. Despite their effectiveness, these approaches can compromise accuracy or be sensitive to weighting strategies, making them challenging to train. In this paper, we deeply analyze the causes and effects of the biases and propose a framework to alleviate biases in recommendation from the perspective of representation distribution, namely Group-Alignment and Global-Uniformity Enhanced Representation Learning for Debiasing Recommendation (AURL). Specifically, we identify two significant problems in the representation distribution of users and items, namely group-discrepancy and global-collapse. These two problems directly lead to biases in the recommendation results. To this end, we propose two simple but effective regularizers in the representation space, respectively named group-alignment and global-uniformity. The goal of group-alignment is to bring the representation distribution of long-tail entities closer to that of popular entities, while global-uniformity aims to preserve the information of entities as much as possible by evenly distributing representations. Our method directly optimizes both the group-alignment and global-uniformity regularization terms to mitigate recommendation biases. Extensive experiments on three real datasets and various recommendation backbones verify the superiority of our proposed framework.

Figures (9)

• Figure 1: We analyze biases in the results of three typical CF models—BPRMF Rendle2009BPRBP, LightGCN He2020LightGCNSA, and SimGCL Yu2021AreGA on the Douban-Book dataset Yao2020SelfsupervisedLF. To facilitate our illustration, we categorize items and users into groups based on their popularity in the training set. We then evaluate, based on the TopK recommendation lists, the recommendation frequency ($Item ~ Recall$) for each item group and the accuracy performance ($NDCG@20$) for each user group.
• Figure 2: Representation distribution of the Douban-Book dataset on $\mathcal{S}^1$. We plot the representation distributions using Gaussian Kernel Density Estimation (KDE) in $\mathbb{R}^2$ and von Mises-Fisher (vMF) KDE on angles (i.e., $\arctan2(y, x)$ for each point $(x, y)$ on $\mathcal{S}^1$). Specifically, we categorize items and users into two groups based on their popularity: blue represents popular items/users, while green denotes unpopular items/users.
• Figure 3: Representation visualization of items and users in the Douban-Book dataset. We randomly selected 500 items and 200 users and utilized T-SNE to visualize the representation spaces of BPRMF and LightGCN, respectively. In the visualization, green dots represent users, while blue and orange dots represent popular and long-tail items, respectively. It is evident that the number of long-tail items significantly exceeds that of popular items.
• Figure 4: An illustration of the AURL framework begins with input data being encoded through CF-based models to map users and items into the representation space. Subsequently, these representations are constrained by two modules: group-alignment $\mathcal{L}{align}$ and global-uniformity $\mathcal{L}{uniform}$, which work together to generate unbiased representations. Finally, the interaction function utilizes these unbiased representations to predict scores for user-item pairs, $s(u,i)$, as part of the recommendation task $\mathcal{L}_{rec}$. It is important to note that while the diagram specifically focuses on item-side debiasing, similar operations are conducted on the user side as well.
• Figure 5: Metrics and performance of BPRMF and LightGCN experiments are visualized. Each point on the plot represents a trained encoder, with its $x$ and $y$ coordinates indicating group-alignment and global-uniformity properties, respectively. The color of each point denotes validation set accuracy, measured by $NDCG@20$, and debiasing metrics, represented by $PRU$. Stars on the plot indicate the points where convergence was achieved.