Side-Informed Attention Collaborative Filtering for Enhanced Wine Rating Predictions

초록

Rating prediction in recommendation systems is a core task that quantifies user preferences and enables personalized recommendations. This study focuses on the wine domain, which is characterized by rich side information, and aims to enhance prediction accuracy by incorporating both user-item interactions and wine-specific attributes such as vintage, grape variety, acidity, and region. However, most existing models primarily rely on user-item interactions, limiting their ability to effectively capture the complex characteristics of wine. To address these challenges, we draw inspiration from the observation that individuals often attend to different characteristics of the same wine during wine tasting. Accordingly, we propose a side-informed attention collaborative filtering (SIACF) architecture that integrates attention mechanisms with matrix factorization. Experimental results demonstrate that the SIACF improves rating prediction accuracy on the X-Wines dataset. Compared to the neural collaborative filtering baseline, SIACF reduces MAE by 2.98% and RMSE by 2.08%, confirming measurable improvements in predictive performance. Additionally, to assess the generalizability of the proposed model, we conducted additional validation on the MovieLens dataset, obtaining competitive results relative to recent studies. The main contributions of this study are as follows: the introduction of a novel approach to integrate rich side information into wine rating prediction; the design of a new model architecture to improve recommendation performance; the demonstration of enhanced prediction accuracy on domain-specific data; and the validation of the generalizability of the model in an external domain. Although the SIACF focuses on the wine domain, future work will explore graph neural network-based models to further enhance generalizability across broader recommendation scenarios.

키워드