Research On Film Personalization Recommendation Based On Resnet&FM Model

The recommendation system is one of the effective means to solve the information overload problem.It can improve the efficiency of users to obtain effective information,eliminate information barriers,and increase the value of information.In recent years,as the abstract unstructured data in the network grows at a geometric speed,the traditional film recommendation system cannot be accurately recommended for specific users in the new data environment,while the deep learning model is layer-bylayer processing and nonlinear.The network structure has a strong ability to learn about abstract unstructured data.Therefore,it is of great significance to introduce deep learning into the traditional film recommendation system.By combing the existing research results,it is found that content-based recommendation and matrix-based decomposition-based recommendation have the defects of abstract user feature and movie item feature extraction difficulty,while the collaborative filtering-based recommendation is sparse when the user and project scoring matrix are sparse.It is difficult to recommend,but it can be effectively solved by introducing a deep learning model.After comprehensive analysis,this paper carries out the following research work according to the model-based collaborative filtering idea:(1)According to the characteristics of the data in the movie recommendation scene,the feature engineering of the film recommendation data is studied.By introducing word embedding and Word2 vec,the dimension explosion caused by One-Hot coding and the difficulty of text feature extraction are solved.(2)A comprehensive analysis of the mathematical principles,advantages and disadvantages of a single shallow learning model and a deep learning model,and improved the Wide&Deep model from the two perspectives of feature fusion and model training.The factorization machine is introduced on the side of the Wide to automatically learn all the second-order display association features,which makes up for the deficiency of the original linear model relying on artificially constructed correlation features;The residual structure is introduced on the Deep side,which alleviates the phenomenon of gradient dissipation of deep neural networks.The Resnet&FM fusion model for the film's personalized recommendation scene was constructed and compared with the Wide&Deep model.The RMSE increased by 2.1% and the MAE increased by 1.3%.(3)Conduct research on Resnet&FM model activation function and weight initialization strategy selection.The deep learning model's ability to learn highdimensional abstract features is mainly determined by the number of nonlinear neurons and the parameters of the whole network.By analyzing the mathematical principles of neurons and parameter initialization strategies,and through experimental comparison,it is proved that based on Kaiming.The Resnet&FM model for initialization and Relu functions works best.Compared with the standard initialization with the Sigmoid function,the RMSE is increased by 3.3% and the MAE is increased by 3.6%.Compared with the Xavier initialization with the Tanh activation function,the RMSE is improved by 4.3% and the MAE is increased by 4.4%.(4)Conduct research on Resnet&FM model activation function and weight initialization strategy selection.The deep learning model's ability to learn highdimensional abstract features is mainly determined by the number of nonlinear neurons and the parameters of the entire network.By analyzing the mathematical principles of neurons and weight initialization strategies,and through experimental comparison,it is proved based on The Resnet&FM model of Kaiming initialization and ReLU functions works best.Compared with the standard initialization with the Sigmoid function,the RMSE is increased by 3.3% and the MAE is increased by 3.6%.Compared with the Xavier initialization with the Tanh activation function,the RMSE is improved by 4.3% and the MAE is increased by 4.4%.