Research And Implementation Of Movie Recommendation System Based On Deep Learning

With the rapid development of the Internet and the exponential growth of data information,it is easy for users to get lost in the massive data,which is "information overload".Especially in the film field,such problems are more prominent.As an effective means to alleviate information overload,recommender systems have been widely studied.At the same time,the recommendation system can increase the stickiness of users to the Internet platform and bring huge profits to enterprises,which has almost become the "standard" of Internet enterprises.It can be seen that the recommendation system has great research value.In this paper,the research is carried out in the context of movie recommendation,and some ideas and algorithms are proposed for several common practical problems in the recommendation scene.The details are as follows:(1)In the recommended recall stage,the DSSM two-tower recall model is usually used to generate an embedding vector(Embedding)for each user and item,and then the nearest neighbor vector search algorithm is used for vector recall.However,when this two-tower model generates the user embedding vector,the user’s basic attribute features and user behavioral features are mixed input,resulting in behavioral features that may cover the user’s basic attribute features,and it is difficult for the embedding vector to express the basic attribute features;in addition,the original two-tower model directly The input of user behavior features ignores the sequence information in the user behavior,which weakens the expressive ability of the user embedding vector.This paper proposes an improved DSSM two-tower recall model,which uses two neural networks to replace the user towers in the original two-tower model to deal with attribute features and behavioral features respectively,so as to avoid the user’s forced behavioral features to cover the attribute features.At the same time,in the neural network processing behavior features,the user interest is modeled,and the attention mechanism is used to combine user behavior and user interest labels to learn the information hidden in the behavior sequence,and further enhance the expressive ability of the user embedding vector.In order to verify the effect,this paper conducts experiments on two public datasets,Movie Lens-1M and Netflix,and compares it with other commonly used recall models.The experiments show that the recall model proposed in this paper has better performance.(2)In the recommendation ranking stage,this paper proposes a new model Deep AFM-LSTM for the shortcomings of the classic recommendation model Deep FM.When the DeepFM model performs feature combination,it does not distinguish the importance of features,which may introduce useless noise,which makes the model converge slowly and reduces the model performance.At the same time,Deep FM treats user behavior in isolation,ignoring the associated information before and after user behavior.The model proposed in this paper introduces an attention mechanism on the basis of Deep FM,which adaptively learns the output weight of the second-order feature combination,avoids the effect of noise on the model caused by invalid features,and accelerates the model convergence.In addition,the LSTM module is added to the model to learn The implicit information in the user behavior sequence can fully mine the data features.Experiments show that the model has good recommendation performance.Finally,combined with the recommendation recall model and recommendation ranking model proposed in this paper,a movie recommendation system with complete functions and good interaction is realized.