TaxoGCN: GCN 기반 추천 시스템 성능 향상을 위한 아이템 분류체계 통합

TaxoGCN: Integrating Item Taxonomy for Enhancing GCN-Based Recommender System Performance

With the rapid expansion of the e-commerce market and the increasing diversity of products and services, the importance of recommender systems has become increasingly significant. Traditional recommender systems have enhanced the recommendation performance through Matrix Factorization (MF) or Graph Convolutional Network (GCN)-based models. In particular, GCN-based models such as LightGCN have achieved superior performance by effectively capturing indirect user-item interactions. However, exclusive reliance on user-item interaction data poses challenges for improving personalization, leading to increased interest in integrating side information such as item taxonomy. Item taxonomy, organized hierarchically from broad categories to fine-grained subcategories and individual items, facilitates learning latent similarities between items. It facilitates the extraction of generalized features at higher levels and detailed features at lower levels, thereby offering a multi-layered representation of user preferences. In this study, we propose TaxoGCN, a novel GCN-based recommender model that integrates item taxonomy. TaxoGCN extends the framework of LightGCN by incorporating user-item interactions along with user-category and item-category relationships, thereby enhancing recommendation performance. Experiment results using real-world data show that TaxoGCN achieves an improvement of 6.7% in Recall@5, 6.4% in Precision@5, 7.4% in NDCG@5, and 8.6% in MAP@5 compared to LightGCN. By explicitly modeling user-item, user-category, and item-category interactions using hierarchical taxonomy within the GCN framework, TaxoGCN effectively captures complex and multi-dimensional user preference patterns, leading to measurable improvements in recommendation accuracy.