Multi-Aspect Graph Representation Feature Integration for Recommender Dialogue System

Recommendation-based dialogue systems aim to capture user preferences via interactive conversations for personalized recommendations. While existing studies focus on modeling user preferences, real-time dialog scenarios face challenges in balancing historical conversation contexts and immediate interests. This study proposes MGIRD, a multi-aspect graph representation approach integrating ordinary graphs and hypergraphs. We use graph structures to model users’ current interests and hypergraphs for historical conversation features, while incorporating historical behaviors in the recommendation module to balance context relevance. A novel item selection mechanism is introduced during dialog generation to naturally integrate recommended items. Experiments on Chinese TG-Redial and English Redial datasets show MGIRD outperforms most state-of-the-art methods in recommendation accuracy and dialog diversity, validating its effectiveness in enhancing recommendation quality and conversational fluency.