Music Playlist Recommendation With Long Short-Term Preference

Music has become more and more important in people’s lives,but it has become a challenge for users to find the right music within a massive amount of music.Therefore,music recommendation is a considerable component of modern music streaming services,which can reduce the users’ choices and improve user experience,and increase business benefits.In this paper,we focus on the music playlist recommendation problem,which aims to provide a suitable playlist for a user by taking into account her interest and music contextual data.To solve this music playlist recommendation problem,our work covers the following several aspects:(1)We propose a data-driven framework,which is comprised of two phases:user/music feature extraction and music playlist recommendation.In the first phase,we adopt a matrix factorization technique to obtain long-term features of users,and utilize the Paragraph Vector(PV)approach,an advanced natural language processing technique,to capture music context features,which are the basis of the subsequent music playlist recommendation.In the second phase,we design two Attention-based Long Short-Term Memory models to achieve a suitable personalized playlist recommendation.We take users’ long-term preferences and music contextual features of songs existing in user’s historical playing records as the input of the model,capture users’ long-term preferences for music,and generate users’ probability vectors of predicted songs to product a top-k recommendation.(2)Based on the data-driven two-segment music recommendation framework,we propose a novel music playlist recommendation framework with users’ long-term and short-term preferences which optimizes the extraction process of users’ long-term preferences.We also propose two Attention-based Bidirectional Long Short-Term Memory models,which can capture users’ long-term preferences for music from users’ favorite playlists.This method can effectively avoid the sparsity problem in the rating matrix and improve the recommendation accuracy.(3)We conduct extensive experiments on a real-world dataset and study the impact of the key parameters.We compare the effectiveness of our proposed solution to the existing solution,verifying the practicability of our proposed methods.