Research And Impletation Of Movie Recommender System Based On Improved Fusion Model

Internet technology has been widely used recent days,and people are overwhelmed by all kinds of information.As a consequence,it's hard for them to find contents that they need,users may run into a lot of problems,which are brought by information overload.People will meet this kind of difficulties especially in some movie website.It is difficult for them to find some movies based on their own interests,which may lead to time and energy consumption.Recommender system solve this problem by sending contents which are interested by different users,making up the shortage of search engines.This kind of technology is receiving more and more attention of scholars and industry.Firstly,we introduce the core theory and implementation of Collaborative filtering(CF),and it can be divided into CF based on neighborhood and CF based on model.CF based on neighborhood calculate similarity through Scored matrix,and making recommendation through neighborhood prediction,but the problem of matrix sparseness,user's interest change and Cold start are hard to solve.CF based on model,such as SVD decomposition and latent factor model,can solve some of the problems by using machine learning method and gain a better prediction accuracy,but they have their own shortage as well.Secondly,We introduce principle and method of singular value decomposition.Aiming at its space and efficiency problem,latent factor model based on the gradient descent are expounded.Moreover,biasedSVD is introduced by adding baseline,and RMSE comparison is made between LFM and biasedSVD.Characters and limitations among LFM and collaborate filtering are compared.Moreover,we use k nearest neighbor model to correct error of biasedSVD.We indicate that the similarity calculation is too simple,and interest change of different users has some shortage.Aiming at the similarity problem,a new comprehensive item similarity based on information entropy is proposed to combine with cosine similarity linearly.Aiming at the problem of interest change,we design an interest declining weight based on the Ebbinghaus memory curve theory and divide all users into different age group to set attenuation coefficient,thereby strengthening the influence of recent behavior to the recommendation.Some experiments were made to test the methods on the Movielens dataset,and encouraging results were obtainedAt last,we develop a movie recommender system based on B/S structure,using improved merging algorithm.The system utilize the JSP and servlet technology to develop the function of personalized recommendation,hot movie recommendation,user rating and movie management,reaching the goal of recommending for different users,and it verify the accuracy of the algorithm.