Research And Implementation Of Movie Recommendation Algorithm Based On Collaborative Filtering

With the popularization of the Internet,the application of e-commerce is also more and more widely.Thus,the data increase rapidly.We must admit we face with such issues as the surge in data volume and the diversification of data structures when we enjoy the convenience that the Internet brings to us.How to make full use of these data to find useful information has become an inevitable challenge.Therefore,the application and development of recommendation system have gradually become the focus of Internet researchers.The conventional collaborative filtering algorithm appeared earlier and was widely used in the recommendation system,such as Taobao,Dangdang,Amazon and other sites.However,the collaborative filtering algorithm still faces cold start problems and scalability problems in the personalized recommendation system.This paper focuses on the cold start problem and scalability problem of the conventional collaborative filtering algorithm.The specific work includes the following three aspects:In order to solve the cold start problem of the conventional collaborative filtering algorithm,this paper proposes using movies'attributes to calculate the similarity.In traditional solutions,random,average or modest recommendations are frequently used to recommend new movie for users.In this paper,we not only used the user's rating information of the movies,but also use the movies' attributes information to calculate similarity between any two movies.According to the meaning of the attributes,we define different ways to calculate similarities so that new movies have a similarity with other movies and ensure they can be recommended to users.Therefore,it effectively alleviates the cold start problem existing in the conventional item-based collaborative filtering algorithm.Design and implement a structure that recommends hybrid models by using BP(Back Propagation)neural network.Traditional predictive models often combine multiple similarities linearly,so they had poor applicability.In this paper,the self-learning's characteristic of BP neural network is used to fuse the similarity calculated by the conventional item-based collaborative filtering algorithm and the similarity calculated by the movie attributes.Then the total similarity is used to predict the movie score of the user.The error between the prediction score and the true score is used to adjust the model to ensure the model's adaptability.In order to solve the scalability problem of the conventional collaborative filtering algorithm,this paper proposes an algorithm for sample sampling.The existing clustering algorithms can not guarantee that each cluster can achieve better results,and the dimension reduction technique not only results in the loss of information,but also can not be guaranteed when the dimension is high.The movie data will not have a lot of changes in the attribute dimension,but the number of evaluation is continuously increasing.Therefore,the sample sampling method is proposed.According to the similarities between the movies,the algorithm mainly uses the idea of TF-IDF and PageRank algorithm to extract the representative movies and uses these representative movies to train the model.In the forecasting phase,the algorithm reduces the prediction space by sampling so that it responds faster to users.The proposed algorithms were compared with other existing algorithms which solve the cold start problem and scalability problem to prove the effectiveness of the proposed algorithms.Experimental results show that AW-CF algorithm proposed in this paper has smaller MAE and RMSE.And sampling AW-CF algorithm not only guarantees the recommended effect,but also takes less time.