Neighborhood-based User Preference Modeling for Conversational Recommender Systems

In this paper, we propose a new end-to-end framework called NUPM (Neighborhood-based User Preference Modeling). In order to enrich the item information and better understand the user intent, we highlight that similar neighbors are essential sources of user preference except for dialogue history in a conversational recommender system. Specifically, we extract sentiment-consistent words and entities, then introduce the neighborhood-based user preference modeling process, performing knowledge-enriched and neighborhood-based recommendations for item suggestions. Finally, we adopt an encoder-decoder based on standard Transformers for response generation. Experiments on the public dataset have shown the effectiveness of our method on both the dialogue system and recommender system.