Movie Recommendation Algorithm Research Based On Deep Factor Network

With the development of society,watching movies on the Internet has become an important source of spiritual entertainment for the people.Movie recommendation systems can bring users a good viewing experience,meet their personalized needs and bring huge advertising revenue for video websites.Movie recommendation system involves many technologies,the movie recommendation algorithm is the core technology,whose purpose is to mine the information accurately that users need or are interested in.The movie recommendation algorithm based on the deep neural network has the characteristics of strong learning ability and real-time performance.It also has the problems of insufficient learning of low-order feature combination and inability to use auxiliary information such as movie title.To solve these problems,the movie recommendation algorithm based on the deep factor network is studied to improve the accuracy of movie rating prediction.The main research contents and achievements are as follows:To solve the problem of how to learn the combination of low-order features in a movie recommendation system to improve the accuracy of movie rating prediction,this thesis analyzes the characteristics of the high-dimensional sparse movie recommendation data and designs a recommendation module based on the factor machine(FM).To improve the learning ability of the second-order feature combinations,the FFM algorithm which is more accurate is introduced.In this thesis,the gradient boosting factor machine(GBFM)algorithm is proposed to learn the second-order cross information of important feature combinations.The gradient boosting decision tree is used to code the samples,and then the coded samples are sent to the FM algorithm for training.The comparison experiment shows that the FFM algorithm and the GBFM algorithm are 6.0% and 5.3% lower than the FM algorithm in the RMSE value of movie rating prediction.To solve the problem of how to learn the combination of high-order features and auxiliary information such as movie title to simulate the real movie recommendation scene,the Deep FM model is improved and the recommendation module based on DNN is designed.The text convolution network is designed to extract the feature of the movie title.The important statistical features are obtained by data analysis,and the original feature set is added.According to the number of categories,different embedding dimensions are adopted.Through the setting of multi-layer hidden layer and incentive function,the learning of high-order feature cross information is realized.Compared with the original Deep FM model,the improved DNN model has a 4.4% reduction in MSE of movie rating prediction.To solve the problem of how to learn the combination of low-order and high-order features in a movie recommendation system at the same time,improve the generalization ability of the model,to further improve the accuracy of movie rating prediction,the deep factor network model is proposed.The weighted average strategy is used to integrate FFM,GBFM and the improved DNN model,which are respectively used to extract the secondorder cross information of features,the second-order cross information of important feature combination and the high-order cross information of features.The experimental results show that the deep factor network model improves the MSE value and RMSE value by 13.6% and 7% respectively compared with the original Deep FM model.The deep factor network model effectively improves the accuracy of movie rating prediction.