Graph-Structured Driven Dual Adaptation for Mitigating Popularity Bias

TL;DR

This work tackles popularity bias in graph-based recommender systems by revealing that deeper GCN layers exacerbate over-smoothing, which reduces the conditional entropy between popular and unpopular item embeddings and weakens supervised alignment. It introduces GSDA, a dual-adaptive framework with Adaptive Hierarchical Supervised Alignment and Adaptive Re-weighting Contrastive Learning that leverage graph-structure signals and real-time popularity distributions to dynamically adjust alignment strength and sample weighting. The authors provide theoretical links between over-smoothing and alignment efficacy, and demonstrate empirically that GSDA consistently improves recommendation accuracy and robustness across three benchmark datasets, while reducing long-tail bias without fixed hyperparameters. The approach offers a practical path to mitigating the rich-get-richer phenomenon in recommender systems, enhancing both fairness and performance in dynamic item distributions.

Abstract

Popularity bias is a common challenge in recommender systems. It often causes unbalanced item recommendation performance and intensifies the Matthew effect. Due to limited user-item interactions, unpopular items are frequently constrained to the embedding neighborhoods of only a few users, leading to representation collapse and weakening the model's generalization. Although existing supervised alignment and reweighting methods can help mitigate this problem, they still face two major limitations: (1) they overlook the inherent variability among different Graph Convolutional Networks (GCNs) layers, which can result in negative gains in deeper layers; (2) they rely heavily on fixed hyperparameters to balance popular and unpopular items, limiting adaptability to diverse data distributions and increasing model complexity. To address these challenges, we propose Graph-Structured Dual Adaptation Framework (GSDA), a dual adaptive framework for mitigating popularity bias in recommendation. Our theoretical analysis shows that supervised alignment in GCNs is hindered by the over-smoothing effect, where the distinction between popular and unpopular items diminishes as layers deepen, reducing the effectiveness of alignment at deeper levels. To overcome this limitation, GSDA integrates a hierarchical adaptive alignment mechanism that counteracts entropy decay across layers together with a distribution-aware contrastive weighting strategy based on the Gini coefficient, enabling the model to adapt its debiasing strength dynamically without relying on fixed hyperparameters. Extensive experiments on three benchmark datasets demonstrate that GSDA effectively alleviates popularity bias while consistently outperforming state-of-the-art methods in recommendation performance.

Figures (7)

• Figure 1: Illustration of representation collapse. The popular item ($i_1$) interacts with multiple users ($u_1,u_2,u_3$), obtaining a generalized representation, while the unpopular item ($i_2$), interacting with only one user ($u_2$), suffers from representation collapse due to insufficient contextual information.
• Figure 2: ((Left) Variation in conditional entropy and embedding similarity across GCN layers, illustrating how deeper layers lead to increased homogenization. (Right) Effect of applying supervised alignment at different GCN depths, highlighting its impact on recommendation performance and the diminishing gains at deeper layers.
• Figure 3: The Graph-Structured Dual Adaptation Framework (GSDA) consists of two key modules: (1) Adaptive Hierarchical Supervised Alignment, which reduces over-smoothing in deeper GCN layers; and (2) Adaptive Re-weighting Contrastive Learning, which dynamically adjusts sample weights based on real-time popularity distributions.
• Figure 4: Comparative Analysis of Recommendation Performance Across Different Item Popularity Groups. This figure highlights the variations in performance metrics, specifically $NDCG@20$ and $HR@20$, across different item popularity groups on the Gowalla dataset, showcasing the impact of popularity bias on recommendation effectiveness.
• Figure 5: Comparison of recommendation performance ($NDCG@20$ and $HR@20$) for varying hyperparameters $\lambda_{1}$ and $\lambda_{2}$ against the optimal baseline on the Gowalla dataset, illustrating the sensitivity and effectiveness of our adaptive modules.