Research On SVD++ Linear Regression Recommendation Algorithms For Stability Problem

In recent years,personalized recommendation system technology has been widely used in e-commerce,advertising sales and other Internet industries.In the absence of clear user needs,the recommendation system can model the user's interests and provide thousands of people with thousands of information and products by analyzing the user's behavior.Aiming at the problems of weak scalability,sparse data and low recommendation accuracy of traditional collaborative filtering algorithm,an improved collaborative filtering recommendation algorithm based on fuzzy partition clustering is proposed in the third chapter.In the traditional cosine similarity calculation method,timediff-item factor,popular-item factor and Nonfashion-item factor are introduced to improve the similarity calculation results,so as to avoid the human objective factors causing the similarity calculation results to deviate from the real situation;at the same time,GIFP-FCM algorithm with improved fuzzy partition is introduced to make the attribute special.In order to test the effectiveness of GIFP-CCF+algorithm,we construct an index matrix for similar items and search for the nearest neighbors of items to form recommendations between the same indexes.In order to test the effectiveness of GIFP-CCF+algorithm,we simulate and compare it with Kmeans-CF,FCM-CF and GIFP-CCF algorithms on Netflix datasets and MovieLens datasets,and prove that GIFP-CCF+algorithm is effective in recommendation aggregation.The results and accuracy of recommendation have certain advantages.GIFP-CCF+recommendation algorithm has the disadvantage of being susceptible to the influence of clustering cluster number,which leads to the unstable recommendation results.In the fourth chapter,a time-effect-based SVD++linear regression recommendation algorithm timeSVD++LR is proposed.The model inherits the characteristics of SVD++model using scoring data prediction,and maps the information fusion implicit feedback from both users and projects into a dimension f implicit semantic space,in which the interaction between users and projects is modeled as an inner product.The score value is explained by describing the characteristics of users and items on each factor,and the time effect attribute is fused.In this time effect attribute,both user bias and item bias are related to users and items to improve the lack of stability of time factor.At the same time,the feature vector X~k is constructed according to the prediction score,and the original training data is used as input of linear regression model.Gradient descent algorithm is used to optimize the final cost function,and the regression parameter vector theta is generated to minimize the value of the cost function.The eigenvector X~k and regression parameter vector theta are introduced into the prediction model,and the prediction score of the test set data is obtained by using the prediction model.The simulation results based on MovieLens dataset show that the accuracy of timeSVD++LR algorithm is significantly improved compared with RSVD,SVD++and timeSVD++algorithm models.Compared with GIFP-CCF+algorithm,timeSVD++LR algorithm has better stability.