ScholarWorks@중앙대학교

초록

Graph Convolutional Networks (GCNs) have substantially advanced recommender systems by effectively modeling high-order user-item interactions. However, GCNs exhibit critical limitations when applied to recommendation datasets with typical long-tail distributions, where they tend to amplify existing popularity bias through message-passing mechanisms. To address this challenge, we propose Debiasing with Norm Adjustment (DNA), a simple yet effective method designed to mitigate degree-based embedding disparities in GCN-based models. DNA normalizes only item embeddings to suppress popularity-induced magnitude differences while preserving valuable user-specific information. Furthermore, DNA introduces a learnable item-specific bias term to restore debiased popularity signals. Notably, the bias term is trained independently of the GCN framework, enabling the model to learn debiased representations without inheriting message-passing artifacts. Our extensive experiments show that DNA improves both the recommendation frequency and performance for low-degree items, as well as overall recommendation quality. In addition, DNA can be integrated into existing graph convolution architectures in a plug-and-play manner, without requiring structural modifications to the underlying models. The code for DNA is available at https://github.com/Chanwoo-Jeong-2000/DNA.

키워드