Research On Music Recommendation Algorithm Based On Deep Learning

Music is an important way of entertainment consumption in people’s daily work and life.With the rapid development of digital music,music streaming media services and mobile Internet,users can enjoy cloud music services anytime and anywhere easily.But the following challenge is: how to find their favorite music in the massive music of cloud services quickly.Music recommendation has become an indispensable part of streaming media music service platform gradually.It allows users to locate their favorite music quickly.It also create huge business value while improving the platform’s user experience and increasing user stickiness.The user’s listening record can be regarded as a sequence behavior.So sequence based music recommendation has gradually become the mainstream music recommendation method in the industry.Users’ music preferences have both long-term stability and short-term volatility.They are also related to music emotion.How to model users’ long-term and short-term interests and hobbies and combine them with music emotion has become an urgent problem to be solved in the industry.This thesis studies the modeling of users’ long-term and short-term preference and music emotion in music recommendation,The thesis also puts forward a music recommendation method based on deep learning and combining users’ listening sequence analysis and music content analysis.The main research contents are as follows:（1）Existing sequential recommendation methods either ignore users’ long-term preferences or the relationship between historical information and the current situation when modeling users’ short-term preferences.These disadvantages result in unsatisfactory recommendation effects.To solve this problem,this thesis proposes a music recommendation model LSTPM based on long-and short-term preference modeling.Firstly,the user’s listening record is segmented into multiple historical sequences and current sequences,which are encoded by multiple Long Short Term Memory networks.Then,for historical series,serial temporalities are proposed to obtain serial representations,and long-term preferences are modeled using nonlocal operations.For the current sequence,the average pooling operation is used to retain the music information,and the user’s short-term preference is obtained.Finally,Softmax function is used to recommend long-and short-term preference.Experiments on two large public datasets,1K-Users and MIGU,show that the model NDCG@10 reaches 0.5449,which is better than the existing model.Ablation experiments further verify the validity of the model.（2）In view of the multiple information characteristics of social media platforms,a hybrid recommendation model AM＿LSTPM is proposed,which integrates music emotional attention and users’ long-and short-term preferences.First of all,using the multi-information characteristics of social media music platform,we learn the emotional characteristics of music from music acoustic signals,lyrics and comments.Then,the key factors of music recommendation are captured by multi-layer attention mechanism by integrating the potential expression of music emotion.Finally,the music attention representation is integrated into the user’s short-term preference and combined with the user’s long-and short-term preference based on context awareness,a music recommendation list is obtained.Experiments on a large real data set prove the validity of the model,and its NDCG@10 reaches 0.5769,which further improves the recommendation performance.In addition,detailed feature contribution analysis is carried out to further verify the model and prove the importance of modeling music emotion and users’ long-and short-term preferences in music recommendation.