Research On Music Recommendation Algorithm Based On Self-Attention Mechanism

With the rapid development of information technology,music works have also embarked on the road of information.Nowadays,more and more people choose to listen to their favorite music on the Internet.At the same time,the Internet is also the carrier of a large number of music works.In the face of such a huge number of music types and Internet,how to quickly find the music people like is particularly important.The music recommendation system can solve this problem well,realize the transformation from finding music by people to finding music by people,improve the transmission efficiency of music,and improve the listening experience of users.In this paper,it is in this technical background and demand to promote the music recommendation algorithm to do in-depth research.Because of the strong relevance of user's conversation behavior in different scenarios,this paper proposes a music recommendation algorithm based on conversation record,which uses self-attention mechanism to extract the characteristics of conversation record.Use the Last.fm Data sets are used to evaluate the performance of the proposed music recommendation model.The main design of the experiment is the setting of model super parameters,the influence of time factor and long-term preference on the recommendation effect,and the comparative analysis with the traditional algorithm.The experimental results show that the algorithm proposed in this paper has good performance in music recommendation.In this paper,based on the user's preference for music in time and the characteristics of being easily affected by time factors,a music recommendation algorithm based on conversation and self-attention mechanism with time factors and long-term preferences is proposed to effectively model music and user preferences,and enhance the ability of short-term feature extraction and long-distance feature acquisition of the model for music listening records.The main research work of this paper includes:firstly,the user's long-term preference is represented by the music he listens to in a certain period of time,and the vector mean corresponding to the music he listens to in this period of time is taken as the initial input of the conversation,which can not only represent the long-term preference,but also alleviate the cold start problem of the conversation;secondly,the multi-layer multi head one-way self-attention is adopted as the feature extraction The extractor can enhance the feature extraction ability,especially for the long-distance feature extraction ability,which is not limited by the time series,and can be calculated in parallel;thirdly,the time factor is incorporated into the model,which is more conducive to capture the changes of user's music preference in a short period of time,and solve the problem of sequence loss of self-attention mechanism at the same time;fourthly,Dr is introduced into the model training process Fifth,in the last output layer,negative sampling technology is used to reduce the bottleneck of model training and improve the training speed.In the verification set,by selecting the appropriate learning rate,batch size,self-attention header number,self-attention layer number and hidden layer dimension,the paper compares and analyzes with the collaborative filtering,session word2vec,item KNN and other recommended algorithms under the general evaluation index.The experimental results show that the model proposed in this paper has good user music preference modeling ability and time the introduction of factors and long-term preferences also has a positive impact on the recommendation effect of the model.