The Research And Implementation Of Music Content Semantics Based Recommendation Algorithm With Deep Learning

With the development of the Internet,recommendation algorithm has been widely used in e-commerce,advertising,community and so on,and the collaborative filtering algorithm is most widely used because of its simple implementation and clear process.However,traditional recommendation algorithm has the problem of long cold start time,Matthew effect and it doesn't make enough use of information in content.Among the existing approaches to cold start issues,the recommended content is not fully utilized.If the content is treated as one of an information source in the recommend system,then we need an appropriate model to learn the "semantics" of the content.Therefore,the purpose of this thesis is how to analyze and model the semantics of music,and finally make use of this information for recommendation.In this thesis,we first introduce the basic principles of deep learning and then analyzes its advantages in content semantic analysis.Secondly,massive music data was collected,clarified and arranged after the analysis of public music data set,and the user data and music data were segmented and integrated.Base on this,we propose a recommendation model based on content semantics named convolutional recurrent neural network recommendation(CRNN).In this model,the audio is segmented and processed into Mel spectrum,and then we use convolution neural network to directly train the Mel spectrum as the network input due to its strong power in learning image.A recurrent neural network was put on the output of the convolutional neural network for learning the feature between different paragraphs of the same track file.Finally,we build a user model by the user's playing data,which include the user's preference on different types of songs,and then recommendation is made according to the similarity between the output features of the CRNN network and the user's preference.In this thesis,regularization and dropout solutions are proposed for over fitting and gradient disappearance when training recurrent neural networks,and successfully solve the problem of over fitting and gradient disappearance of complex models.Finally,we present the results and analysis of the comparative experiments of both classifier and recommendation algorithm respectively.The results show that the model presented in this thesis outperforms traditional algorithms on multiple test datasets used in experiment and also in the independent cold-start recommendation algorithm test.The cold start of the traditional recommendation algorithm problem has improved in this model.Result of these experiment shows that the proposed model is effective and feasible.