Work

A Gated Review Attention Framework for Topics in Graph-Based Recommenders

Year: 2025
Type: conference-paper
Abstract: Recommender systems significantly reduce information overload by curating personalized content on digital platforms, thereby enhancing user experience. Traditional models often rely on sparse rating data, overlooking the rich semantic signals embedded in user reviews. To address this, we propose the Gated Review Attention Framework for Topics, a novel graph-based recommender system that integrates review-derived topic information with user-item interactions. Our approach leverages BERTopic to extract interpretable semantic topics from textual reviews and then embeds them into a graph attention network architecture. A gating mechanism dynamically regulates the influence of these topic representations relative to latent user and item embeddings, enabling adaptive feature fusion. We evaluate GRAFT on three benchmark datasets: Amazon Movies and TV, IMDb, and Rotten Tomatoes. Comparing it against classical and neural baselines, including SVD, DeepCoNN, and KANN. Experimental results demonstrate that GRAFT consistently achieves the lowest RMSE across all datasets, indicating superior rating prediction accuracy. Although traditional models perform better on ranking metrics, GRAFT achieves superior accuracy (lower RMSE), and our qualitative analysis demonstrates more substantial semantic alignment.
Language: English
Open Access status: hybrid
OpenAlex ID: W4416744729
FWCI: 3.676
Cites: 26
Cited by: 7
Related to: 10