Research On Session-Based Recommendation System With Deep Sequence Model

With the rapid development of information technology,recommendation system has been widely used in various e-commerce platforms and streaming media websites.Most of the existing recommendation systems give personalized recommendation schemes based on user identification and long-term historical behaviors.However,due to many real application field merging,users' long-term behavior data and identity can not be obtained.In this case,the traditional recommendation algorithm will no longer be applicable.The corresponding recommendation results are given by clicking sequence of users in the current session.This recommendation method is also called session recommendation system.Most of the existing model of session recommendation system is based on one-way sequential neural network model,which has the following problems: 1)the phenomenon of "gradient disappearance" will appear with the increase of network layer,which makes it difficult to capture the dependence in long-term sessions;2)The structure of serial network of one-way sequence model limits the computing speed of the model;3)One way sequence model is built by using left to right information.This modeling method can not effectively capture the dynamic changes of user interest.In view of the above problems,the main work of this thesis is as follows:1.this thesis proposes a session recommendation system model based on two-way sequence.It can extract the interest of user session sequence information through single layer attention mechanism,and solve the defects of one-way sequence model in the long sequence "gradient disappearance",serial calculation and one-way sequence neural network learning user behavior sequence.The model also extracts the user dependence in long-term sessions by using attention mechanism,and captures the dynamic changes of user interest preferences through the two-way long and short-term memory network model.The experimental results show that the model has better recommendation effect.2.this thesis proposes a deep two-way sequence conversational recommendation system model,which is realized by stacking multi-layer attention network,which solves the problem that the shallow attention network can not effectively capture the changes of users' deep interest.Based on the deep bidirectional sequence model,two methods of parameter reduction are proposed.The first method is factorization parameterization,which decomposes the large vector factorization into two small matrices in the embedded layer.The second method is to share parameters across layers,and to reduce the amount of parameters in the whole connection layer and attention layer.The validity of the model is verified by experiments on three common data sets.3.this thesis designs a personalized recommendation system for movies.In the design of personalized recommendation system of movies,this thesis adopts the open source bootstrap framework and Django framework to design,and realizes the data processing module,cache module and recommendation module.Finally,the thesis tests the function modules of the system,and the functional modules are running normally,which verifies the reliability of the modules.