Research On Recommendation Algorithms Based On Deep Learning And User Behavior Sequences

In recent years,with the rapid development of e-commerce,online shopping is playing an increasingly important role in people's daily lives.However,with the increasing number of products in e-commerce websites,users find it more and more difficult to find out the products they are interested in.Therefore,how to design recommendation systems to help users select products of interest from a large number of products has gradually become a popular research direction in academia and industry.Traditional recommendation systems that use a static approach to model user interaction information can only capture the static interest preferences of users.In contrast,a recommendation system based on sequential modeling of user behavior is able to capture dynamic interest preferences of users.In order to improve the accuracy of sequential recommendation,this thesis firstly proposes an innovative multi-pre-training sequential recommendation model by pre-training the input with reference to the pre-training task in natural language processing.Secondly,in order to enrich the diversity and variability of the pre-training inputs,an attention model is introduced to learn the important information from multiple pre-training models.Finally,based on the deep neural network knowledge to model user behavior sequences,the sequential recommendation model is proposed PMCA-Bi LSTM.the main work of this thesis is as follows:(1)To address the problem of sparse data in recommendation systems,pre-training of inputs based on models such as Word2 Vec,Glove and Pro NE is proposed.By pre-training the input information,we are able to learn the multi-hop relationship of user nodes and global information,which allows the model to learn richer and higher-order information and alleviate the input sparsity problem to some extent.(2)For the problem of learning users' differentiated interests,attention mechanism is used to fuse multiple pre-training models so that the embedding vector of items carries more a priori information.At the same time,the attention mechanism is also introduced in the user sequence to learn the relationship of multiple interests of different users to solve the problem of model learning the differentiated interests of users.(3)To address the problem of insufficient information in mining user's historical behavior sequences,this thesis firstly adopts max pooling and average pooling schemes to capture users' long-and short-term interests.Secondly,a multi-scale convolutional residual neural network is used to enable the model to learn the global and local information of user sequences; a bidirectional recurrent neural network is used to replace the unidirectional recurrent neural network to better capture the bidirectional dependencies in user behavior sequences.Finally,this thesis uses a residual MLP network to prevent model overfitting and improve model recommendation accuracy without losing valid information.Finally this thesis does a comparison experiment on Amazon dataset.The experimental results show that the recommendation model based on user behavior sequence data proposed in this thesis is improved in recommendation accuracy,NDCG and other metrics compared with popular recommendation,collaborative filtering,DNN and graph neural network models.