Conversational Music Recommendation based on Bandits

Abstract

Music is one of the most popular products in the recommender system, and there have been many various methods of exploring music recommendations. Traditional music recommendations commonly collect users’ feedbacks in limited ways for preference analysis. The text dialogue is a direct and natural interactive mode, providing diversified information. In this paper, we discuss the music recommendation in an innovative scenario – a conversational music recommendation model, which integrates the advantages both from the recommender system and dialog system. This paper adopts a ”user ask, system respond” interactive way to obtain users’ real-time requirements, and users are allowed to express their requirements on music in free text. In order to face the fast-changing music preferences, this paper adopts the bandit-based algorithm to absorb users’ attitudes to the current recommendation, and the results show these methods achieve better performance than baselines. Besides, it also constructs a music-domain knowledge graph to support the richer users’ musical expressions with millions of music items and tens of millions of relations.