Research On Personalized Recommendation Algorithm Based On User Behavior Analysis

With the rapid development of the Internet,the types and scale of data in the network have grown rapidly,and it is difficult for users to discover items of interest.In order to solve the problem of ”information overload”,personalized recommendation algorithms are widely studied and used.The high degree of data sparseness and large data size pose a challenge to the accuracy and recommendation efficiency of the recommendation algorithm.Research on efficient and accurate recommendation algorithms is of great significance for improving user experience and corporate profitability.From the perspective of user behavior,this thesis proposes corresponding solutions to the problems of data sparsity,recommender system scalability,and neighbor search efficiency based on traditional collaborative filtering algorithms:1)This thesis proposes a user similarity model based on multi-dimensional feature Fusion called FMF.The recommendation accuracy rate is low due to the use of a singledimensional measurement model on a highly sparse data set,this thesis calculates the user similarity based on multiple time dimensions and geographic dimensions.And introduce the time attenuation factor to improve the Pearson correlation coefficient,and finally merge the similarity based on time,geography and rating dimensions to obtain a new user similarity measurement model.Experiments show that the collaborative filtering algorithm based on the FMF model has lower prediction errors than the benchmark algorithm on the data sets ml-100 k and ml-1m.2)This thesis proposes a collaborative filtering algorithm based on improved local sensi-tive hashing called O-LSH-CF.Aiming at the problems of low search efficiency and poor system scalability due to the large data size and high dimensionality when calculating neighbors using traditional collaborative filtering algorithms.In this thesis,LSH is introduced to hash the dimensionality of massive high-dimensional user data,and the OPTICS algorithm is introduced for pre-clustering and noise removal,and then performing nearest neighbor search in the corresponding sub-clusters.While hashing the original data to reduce the dimensionality,the search range is narrowed,thereby improving the search efficiency.Experiments show that the operating efficiency and accuracy of the O-LSH-CF algorithm on the data set ml-100 k are higher than the benchmark algorithms.3)This thesis proposes a matrix factorization recommendation algorithm called OTMLMF based on multi-dimensional feature constraints of users and item neighbors.Based on the previous two algorithm models,a matrix factorization algorithm based on user and item neighbor constraints is proposed to alleviate the cold start and sparsity problems of the recommendation system When calculating the neighbors of users and items,the O-LSH algorithm is used to search for neighbors to improve the efficiency of neighbor search,and the user similarity is calculated based on the FMF model.In addition,the user herd effect factor is introduced to measure the confidence of users' ratings,which makes the model more accurately reflect the users' real interests and preferences.Experiments show that the prediction errors of the OTML-MF algorithm on the data sets ml-100 k and ml-1m are lower than those of the benchmark algorithms.Through the above researches,the proposed multi-dimensional feature fusion user similarity model can effectively alleviate the impact of data sparsity on the accuracy of recommendation.The proposed nearest neighbor search model based on improved LSH can effectively improve the nearest neighbor search efficiency and recommendation accuracy of the recommendation algorithm.The proposed matrix factorization recommendation algorithm based on nearest neighbors and multi-dimensional feature constraints can significantly improve the recommendation accuracy of the algorithm.