Users’ Preference-Aware Music Recommendation with Contrastive Learning

Abstract

Music recommendation serves as a critical branch of recommender systems, which is pivotal to capture users’ preference via their historical listening sequence for improving user experience. However, existing efforts in music domain acquire users’ preference in a supervised manner, thus inevitably suffering from the problem of data sparsity due to rare interactions between users and music pieces and further failing to precisely model users’ preference. Inspired by the recent success of self-supervised learning, in this paper, we propose a novel Users’ Preference-aware Music recommendation with Contrastive Learning (UPMCL) method to mitigate the above issues. To be specific, the proposed approach UPMCL first encodes the information of music pieces according to original and augmented listening sequences. Moreover, it employs the contrastive learning to maximize the agreement between mask- and permute-based augmented listening sequences to learn the representations of music pieces. Eventually, the attention mechanism is utilized to integrate different types of users’ preferences to generate the comprehensive users’ preference and further achieve accurate music recommendation. Extensive experimental results conducted on three real-world music datasets clearly demonstrate that UPMCL has the capability in effectively recommending appropriate music pieces to users.