Music Recommendation Technology And Application Research Of Listening Sequence Word Embedding

With the rapid development of Internet and the strong rise of streaming media music service,people can listen to massive music anytime and anywhere.While people are enjoying the digital dividends brought by information growth,they are also facing increasingly serious information overload issues.For this reason,the recommended technology came into being.The listening sequence contains rich music context information,which can effectively describe the user's potential interest preferences.It is a research hotspot in the field of recommendation system at present to analyze the existing recommendation technology and study how to utilize the new generation recommendation technology represented by machine learning to dig out users' short-term dynamic preferences from specific scenes and local sequences and to construct real-time recommendation services that meet users' needs and improve platform benefits.Taking music recommendation as the research object,this thesis proposes a hybrid music recommendation model E-GRU(Embedded-Gate Recurrence Unit,E-GRU)which combines distributed word vector representation(Word Embedding)and gate recurrence unit(GRU)to maximize recommendation accuracy while mining users' potential interest preferences.This thesis can be divided into two parts.The first part is based on the characteristics of strong correlation between the music in the local sequence,segmenting the user's historical listening sequence to form a conversation record.The Word2Vec framework based on Continuous Bag of-Words Model(CBOW)is used to learn the vector representation of music word at the conversation level to realize the coarse-grained extraction of music features;Secondly,the rationality of the model is verified by theoretical derivation of the constructed word embedding model;Finally,the constructed model is experimentally trained on the Last.fm real dataset to prepare input data for subsequent models.The second part combines users' short-term dynamic intentions and long-term stable preferences and proposes GRU music recommendation model E-GRU based on word embedding The addition of GRU makes the model not only enhance the modeling ability for time series information but also dig out the deep music abstract features contained in the sequence and overcomes the problem that the traditional cyclic neural network loses the learning ability of long-distance dependence as the depth of the model increases.The model uses the music word vector instead of the one-hot vector to represent the input data and uses the mean of the music word vector as the initial input to characterize the user's long-term preference.This solution not only fills the lack of shortterm historical information of the short-term preference model,but also reduces the computational cost of the model;Secondly,by using Dropout technology to overcome the over-fitting problem in the model training process,effectively improving the generalization ability of the model,and finally experimentally training the constructed model on the real data set and comparing it with various recommendation algorithms.The experimental results show that the improved technical research scheme and experimental model constructed in this thesis can make more accurate music recommendations.