Research On Some Key Technologies Of User-based Collaborative Filtering Recommendation Algorithm

With the rapid development of Internet and e-commerce,though they enriches people's lives,the problem of information overload has also come with them.Recommender system which provides users with accurate,intelligent and personalized recommendation services is a technology to solve this problem.Determining the number of K nearest neighbors and predicting user's rating on the item are two key steps of the recommended technique.First of all,the number of K nearest neighbors,i.e.the number of similar users,is generally determined by experience or multiple experiments.Therefore,the problem of existing methods like strong subjectivity and cumbersome process affect the accuracy of the recommended algorithm.Secondly,the prediction of ratings: when there are several neighbors,if we use Cosine,Pearson and other classical similarity,the similarity value between users will be 1 mostly.At this point the results calculated by traditional prediction method are mostly similar to the user's average rating value.When there is only one neighbor rating the target item,the contribution of similarity between users is zero to the final prediction score.All the prediction results are the average rating value of target users,and it can't distinguish user preferences well.To solve the above problems,this paper carries out an in-depth analysis and research for the selection of nearest neighbor and the method of rating prediction,and then establishes a K nearest neighbor optimization model and proposes an improved rating prediction method based on the user collaborative filtering algorithm respectively.The main contents can be summarized as follows:(1)Nearest Neighbor Optimization Method Based on Differential EvolutionAlgorithmFirst of all,this paper establishing the optimization model with the minimization of the mean absolute error(MAE)as the objective function combines the user's actual scoring and predicting scores.Secondly,the optimal result is obtained by the differential algorithm.Finally,the superiority of the new method is verified by three metrics,i.e.MAE,precision and recall.The new method breaking the limit of similarity threshold set artificially in traditional K nearest neighbor can quickly find the optimal value of K by differential algorithm.(2)Improved Prediction Method Based on SlopeOne AlgorithmThe method learning from SlopeOne algorithm based on traditional rating prediction methods takes full account of common ratings of the current users and its nearest neighbors simultaneously,reflects the contribution of different nearest neighbor users to the prediction of current users' rating behavior through fusing the similarity,and design an improved rating value prediction algorithm.The new method can effectively solve these problems that the traditional rating prediction method can't distinguish user's preferences effectively,doesn't make full use of the user rating information,and treats the nearest neighbor users identically.In this paper,considering two scenes as cold start and non-cold start,we verify the performance of the two proposed new methods on Movielens,Epinions and Netflix classical data sets.The new methods have obvious advantages over the traditional prediction methods in MAE,precision and recall,and significantly improves the accuracy and recommendation quality of user-based collaborative filtering recommendation algorithm.The two methods proposed in this paper can be applied to the cold start and noncold start environments are well integrated with the existing recommendation system and have high application value.