Applicaton Research Of LSTM Network Based On Word Embedding In Music Recommendation

The data-driven and machine learning-centered recommendation technology can effectively improve the accuracy of information services and enhance users' experience.It has been a research hotspot of information processing technology in recent years.The dissertation takes the typical application music recommendation in media recommendation as the research object,and applies the distributed word embedding representation commonly used in machine learning and design LSTM neural network model to the recommendation algorithm to effectively improve the recommendation quality and efficiency.The main research contents of the thesis include two aspects.One is the distributed music vector learning based on music session record used Word2 Vec,The second is to model the LSTM music recommendation algorithm combined with long-term preference.In many recommended scenarios,users made a series of conversational behaviors in a short period of time can better reflect their psychological state and needs at the time and prossess strong correlation.In this dissertation,propose a music recommendation based on session used Word2 Vec,and Word2 Vec is text vector neural network learning model most commonly used.The dissertation uses Word2 Vec framework to generate music distributed vector representation in session records,in order to effectively reduce the dimension of music vector representation and obtain correlation between music in similar scenes and realize effective extraction of music features.The main research work in this part of the thesis used Word2 Vec and the feasibility theory derivation of music embedding extraction based on conditional probability.The dissertation uses the Last.fm real dataset to experimentally train the model.The main design of the experiment is setting window value and selecting music vector dimension and visualization of experimental results.The experimental results show that the similar 'semantic' songs are closely connected in twodimensional space,which shows that the acquired music embedding is reasonable and effectiveness.Based on the short-term coherence and long-term stability of music hobbies.the dissertation proposes combining long-term preference music recommendation modeling used LSTM which can effectively solve the problem of loss of remote information learning ability when the RNN time span is increased.The research work in this part of the thesis mainly includes: Firstly,The average value of the historical vector is used as the initial input,and the recommendation accuracy is further improved by combining the long-term preference with the short-term preference and solve the problem of session cold-start.Secondly,unify the length of the session sequence facilitate parallel operation.At the same time,the dynamic update idea is used to solve the problem of error increase caused by zero padding.Thirdly,in the model music embedding layer uses Dropout technology to further improve the generalization ability of the model.Fourthly,define a suitable loss function for the personalized recommendation algorithm.whether to recommend.Through several experiments on the verification set,select the appropriate number of hidden layer nodes,learning rate,batch size and other hyper parameters.Based on the general recommendation index system contrast with other session recommendation algorithm,like as session-Word2 Vec,WLSTM.The results show that the proposed model has good dynamic intent extraction ability and can alleviate the session cold-start to a certain extent.