Music Recommendation Research Based On Double-layer Attention Mechanism

In the era of big data,the rapid growth of information on the Internet has aggravated the phenomenon of "excessive information ".The recommendation system can establish a personalized interest model for users by analyzing their past behavioral characteristics and interest preferences,and help the user find the information quickly and accurately.Music is an important way for human beings to express their emotions.It is an art that entrusts the emotions of life.In modern society,it is easier for people to listen to music as a daily way of entertainment.It is suitable to introduce recommendation system in music domain.The accurate music recommendation can not only enhance the user experience,but also bring traffic to the music website to better create business value.Therefore,personalized music recommendation that suits the user’s taste becomes more and more important.The main methods of common music recommendation system include content-based recommendation,collaborative filtering recommendation and mixed method recommendation,etc.Although the existing recommendation algorithms have achieved relatively good results in achieving score prediction and music list prediction,there is still a large room for improvement.For example,in the recommendation algorithm based on collaborative filtering type,the user interest model constructed by generating user-score matrix is prone to data sparsity.In addition,users’ interests and preferences in the near future cannot be obtained only according to the information of users and the song itself,resulting in poor real-time performance of recommendation results.Recommendations for ever the problem of data sparseness and real-time of the this paper presents a music recommendation model based on double attention mechanism,respectively from the feature level and project level access to the user’s interest in music features and music projects of preference,but also in the use of natural language processing text convolution neural network and Word2Vec technology,process and reduces the complexity of the calculation of characteristics.This paper mainly does the following work:(1)The deep learning method can learn the essential characteristics of data.This paper combines music recommendation with deep learning,and proposes a music recommendation method based on double-layer attention mechanism,with various feature information of users and music and user history music listening list.As input,the feature is extracted through the embedded layer of the neural network,and then a two-layer attention mechanism is designed to learn the preference between each feature of the user and the music from the feature level and the item level,and the user’s preference for each music in the music listening list.(2)For text data such as music names and music labels,it is converted into distributed word vectors by Word2Vec technology in natural language processing,and input into text convolutional neural network for feature processing,thereby reducing the amount of calculation and improving music recommendation.(3)For the user short-term interest preference model learned by the double-layer attention mechanism,the vector cosine similarity method is used to calculate the similarity between the user’s interest vector and the music vector,and the music with higher similarity is obtained,and the TOP-N is selected.Music generates a personalized list of music recommendations for the user.By running the algorithm on the Last.fm 360K users dataset and comparing it with the project-based collaborative filtering algorithm,the neural network-based collaborative filtering algorithm,and the single-layer attention mechanism-based method,the proposed method in this paper has significantly improved the hit ranking and the normalized discount cumulative gain index,which proves that the algorithm has better recommendation effect and is also helpful to solve the problem of data sparseness and real-time.