Research And Implementation Of Film Recommendation System Based On Deep Learning And Behavior Sequence

In recent years,the rapid development of Internet technology has led to information overload.At the same time,users’ needs are becoming more and more diverse and personalized.As one of the information filtering technologies,recommendation system is faced with the problem of relying too much on explicit feedback from users,while ignoring the impact of cold start on user retention.Therefore,this thesis studies and proposes a recommendation algorithm based on user behavior sequence and a cold start movie recommendation algorithm for new users.The specific research content of this thesis are as follows:(1)The background and research significance of this thesis are expounded,and the shortcomings of traditional recommendation algorithms,including collaborative filtering recommendation algorithms,are analyzed.At the same time,the research status at home and abroad,advantages and disadvantages of various recommendation algorithms based on deep learning and behavior sequence are studied.(2)In the practical recommendation application,with the passage of time,the interests of users will change accordingly,and the closer the historical behavior to the current moment,the greater the influence on the interest state of the current moment.However,the existing GRU model can only infer the final interest through historical behaviors,without taking into account the characteristic that the influence of historical behavior on the final interest decays with time.An improved model first proposed in this thesis,i.e.,TGRU,combines the time weight with the GRU model to obtain a more accurate expression of interest.Furthermore,the AUGRU model is used to capture the evolution process of those interests related to the final interest,thereby predicting the users’ future interest.Based on the above results and analysis,this thesis finally proposes a movie recommendation model,i.e.,MRTUB,which integrates time weights and user behavior sequences.The experimental results show that the MRTUB movie recommendation model proposed in this thesis has better effect than various superposition models of GRU.(3)To solve the problem that the current cold-start recommendation relies too much on the rating data of users,this thesis focuses on the input features of users and movie features,uses deep learning to calculate the similarity of users,and fuse the features of users and movies.Therefore,a cold-start movie recommendation model RM-SC-MLP based on similarity calculation and MLP network is proposed.The model uses user feature information and MLP to calculate the overall similarity between users,and then gets the candidate movie set of new users according to the interactive movies of similar users.Then based on deep learning,user features and movie features are extracted and fused,and finally the prediction score is obtained.Experimental results show that the cold start movie recommendation algorithm proposed in this thesis is more accurate than other cold start methods.(4)Using the two recommendation algorithms proposed in this thesis,combined with different recommendation scenarios,and using Vue and Spring Boot framework,a movie recommendation system is designed and implemented.