Cross-Domain Explainable Recommendation Using Graph Convolutional Networks and a Topic Model
Kunyoung Kim, Mye M. Sohn, Jongmo Kim · IEEE Access · 2025
Graph Convolutional Network (GCN)-based recommendation systems (RSs) have recently gained popularity for their ability to improve recommendation accuracy by utilizing neighborhood information in user-item interaction graphs. Despite their success, GCN-based systems still face two major challenges, which are the lack of explainability and limited applicability in multi-domain environments. To address these limitations, we propose a novel GCN and Review text-based cross-domain ExplainAble recommendaTion (GREAT) framework that enhances both the accuracy and explainability of cross-domain recommendations. Instead of relying on structured knowledge graphs, GREAT utilizes user-generated review texts to incorporate users’ sentiments and opinions directly into recommendations and explanations. GREAT extracts domain-wide topics using a proposed Term-Weighted Latent Dirichlet Allocation (TW-LDA) model, which reduces domain biases by adjusting for differences in term usage across domains. For recommendation, GREAT introduces TACO (Topic-Aided Cross-dOmain Recommendation), a GCN-based collaborative filtering model that integrates domain-wide topic information to enable effective knowledge transfer across domains. For explanation, GREAT identifies topic-level connections between recommended items and a user’s historical interactions. Experiments on a real-world Amazon dataset spanning three domains demonstrate that GREAT outperforms state-of-the-art baselines in recommendation performance. Additionally, we validate the explainability of GREAT through a case study based on actual recommendation results.