Research And Application Of Recommendation Algorithm Based On Deep Learning And Matrix Factorization

With the prosperity of the Internet industry,people are facing the problem of massive data can not be used effectively and information overload.In order to facilitate users to find the information they want,the recommendation system is proposed as a solution.Among the recommended algorithms,matrix decomposition algorithm is widely used.The implementation of the algorithm only predicts the score based on the existing score data.As the score data becomes sparse,the degree of under fitting of the algorithm becomes higher and higher,which affects the final prediction effect.Aiming at the problems faced by the traditional matrix decomposition algorithm,this thesis completes the following work:(1)The implementation process of the traditional matrix decomposition algorithm only carries out linear interaction between the user and the project hidden vector after decomposition,so as to predict the score.The simple linear interaction causes the problem of large prediction error.To solve this problem,the linear and nonlinear interaction information between users and project hidden vectors is fused by combining the generalized matrix decomposition model and neural network collaborative filtering model.The fused model is named GNMF model.Then,by comparing the GNMF model with the two models before fusion,it is found that the effect of GNMF model after fusing linear and nonlinear interactive information has been greatly improved by comparing the loss function curve and various evaluation indexes.(2)Although the effect of GNMF model is better than that before fusion,the problem of under fitting is still prominent.Based on the GNMF model and combined with the relevant knowledge of the combination model in the deep learning recommendation model,a combination model integrating the generalized matrix decomposition module,neural network collaborative filtering module,finite order feature crossover module and deep neural network module is proposed,which is named DNGFC model.Based on the original GNMF model,DNGFC model is used to capture different types of feature cross information between user and project features by adding finite order feature cross module and deep neural network module as a supplement to improve the prediction effect of the model.The finite order feature crossover module uses Cross Network,Factorization Machine and Compressed Interaction Network for three times.The public movie dataset Movie Lens100 k is used to verify the DNGFC model using different finite order feature crossover modules through ten-fold cross validation.Accuracy,F1 score and AUC are used as offline evaluation indexes to compare the effects between the models.Finally,it is found that the DNGFC model combined with compressed crossover network performs best,The performance of three different indicators is significantly better than other comparison models.(3)A movie recommender system based on Django is designed and implemented,and the best DNGFC model is applied to verify the engineering applicability of the model.