A Music Recommendation Model Based On Neural Network

Over the past decade,the global recorded music market has grown steadily,and streaming music platforms have become a key driver of overall growth.With the rapid increase in the number of music creation and the decline in the difficulty of content acquisition,many music fans are facing the problem of music selection.How to recommend music that meets their preferences for music fans and establish an effective music recommendation system has become the core issue of streaming music platform.In this paper,for the short-term music recommendation problem,based on the hybrid neural network algorithm,an audio content-based music recommendation model is established to mine the temporal dependence and internal correlation of the music in the user’s listening list,so as to explore the user’s current situation and emotional state,and recommend the next appropriate music for the user.The main research contents of this paper are as follows:(1)By extracting the cepstrum feature,time domain feature and energy feature of audio information,using embedded technology to splice them with the structured features of audio to form a song feature matrix,and then extracting a song one-dimensional feature vector containing rich feature information through convolutional neural network;(2)Based on two recurrent neural network models with gating structure and self-attention mechanism,two music recommendation models,BGRU-SA and BSRU-SA,are established to mine the temporal dependencies and autocorrelation in users’ listening lists,and the recall and average reciprocal ranking are used to evaluate the performance of the models.The experimental results show that the recommendation effect of the two models is significantly higher than that of the traditional model and the original model without self-attention mechanism;(3)By comparing the effects of the two improved models on multiple data sets,the analysis results show that the BGRU-SA model has a higher prediction success rate but a slower training speed,which is suitable for scenarios with low data quality or low requirements for training speed,while the BSRU-SA model has a faster training speed but relies heavily on data.It is suitable for scenarios with high data quality or high training speed requirements.The music recommendation model designed in this paper uses the improved model based on a variety of neural network types to analyze the content information of the songs in the user’s listening list,explore the user’s situation and emotional state at the moment,and recommend the next song for him,which achieves a good recommendation effect and has a certain application value.At the same time,the potential of recurrent neural network with gating structure in audio information feature training is explored,and it is found that it has a good performance after introducing the self-attention mechanism,which provides a reference for further research.