Research On Personalized Recommendation Algorithm Based On Data Mining

In this information explosion age,the Internet produces a huge amount of data everyday,how to be able to accurately and quickly find the information they want is an urgent problem to be solved.Recommendation system can help users to find the really valuable information in the immense resources,save the user's search cost,and also improve the user's loyalty to the website,and increase the revenue of the website.At present,Recommendation system is widely used in e-commerce,advertising,social network,digital library,movie and music on demand.The future of social networking sites will be driven by the recommendation system.So the recommended system is widely concerned by academia and the industrial circles.The ACM took the recommendation system as a research subject.The domestic and foreign journals have also done this.Among them,the recommendation algorithm plays an important role in the recommendation system.Therefore,it is necessary to study a kind of efficient personalized recommendation algorithm,to make full use of existing data set,to explore the user deeply,and establish prediction model to predict each user interest degree of items,then for each user formed their own list of recommended,personalized recommendations.This paper research on Personalized Recommendation Algorithm from the following three aspects:(i)Research on the optimization method based on the Latent Factor ModelLFM can identify the hidden themes or categories,and establish the relationship between the features by the implicit theme or classification.The core idea of this algorithm is to establish the user-item rating matrix,and solve the two low dimensional matrix,so that the two low dimensional matrix multiplication can approximately express this rating matrix.The general solution of LFM using gradient descent method to minimize the cost function,but considering the application in the field of recommended systems to deal with massive data,so in this chapter,we propose a parallel method of stochastic gradient descent based on CUDA,to improve the efficiency of recommendation.(ii)Parallel PersonalRank algorithm based on CUDAThe PersonalRank algorithm can be better explained by the random walk theory,but the time complexity of the algorithm has obvious disadvantages.Because each user to recommend items are needed to iterative repeatedly in a bipartite graph,until each vertex of the graph converge.The time complexity is very high,not only cannot provide online realtime recommendation,but also is time-consuming to generate the offline recommended results.Therefore,in order to work around this problem,this paper uses CUDA to parallel PersonalRank algorithm,so as to improve the efficiency of the recommendation.(iii)Hybrid algorithm of Latent Factor Model and PersonalRankIn the application of the recommendation algorithm,outstanding personalized recommendation method is generally fusion on multiple or even a hundred model in order to achieve good results,hence multi model fusion has a vital role to improve the accuracy of the recommendation.Therefore,we propose a hybrid recommendation algorithm fused latent factor model and PersonalRank algorithm based on a random walk.Firstly,using latent semantic predicts rating for user not evaluated items.Secondly,filling user ratings matrix when predict the score greater than a certain threshold,then establishing diagram model of users and items.Thirdly,computing PR value of each user to each item using the PersonalRank algorithm,and top-N recommendation.Finally,in the MoiveLens data set,we validated the predictive rating effect utilizing latent factor model to fill sparse matrix.