Mitigating Popularity Bias in GNN-based Collaborative Filtering through Debiasing Techniques

Graph neural networks (GNNs) have shown remarkable effectiveness in collaborative filtering (CF) by leveraging user-item interaction graphs. However, these models often face challenges due to popularity bias, where the prominence of frequently interacted items overshadows those that are less popular. This bias is exacerbated by the nature of message passing in GNNs, which tends to amplify signals from popular items while diminishing the visibility of niche options. Traditional methods aimed at debiasing, such as re-weighting strategies and post-processing techniques, fall short in GNN contexts as they do not adequately address the bias introduced during the aggregation phase. To tackle this issue, we introduce the Debiasing Popularity Amplification in Aggregation (DPAA) framework. This innovative approach incorporates adaptive interaction weighting that is sensitive to user representations and applies layer-wise weighting to enhance the visibility of less common items. By utilizing a dynamic transition between pre-trained and evolving embeddings, DPAA effectively stabilizes the interaction weights. Our experiments demonstrate that DPAA significantly outperforms existing methods for correcting popularity bias in GNN-based collaborative filtering, showcasing its potential to improve recommendation diversity.