Research On Recommendation Algorithm Based On Clustering And Weighted Matrix Approximation

The emergence and popularity of the Internet has brought more accessible information to users,meeting the demand for information in the era of big data.However,with the rapid development of the Internet,information has exploded and expanded,and the structure of information has also changed.Increasingly complex,the problem of "information overload" is formed,the difficulty and cost of acquiring content and products of interest to ordinary users are correspondingly improved.The increase in the amount of information,on the contrary,has an adverse effect on the utilization of information.The personalized recommendation system is considered to be one of the most effective tools and methods to solve the above problems.Since its introduction,it has been rapidly developed,and it can more accurately select the products and contents that meet the actual needs of users from the dynamically changing information flow.Thus to some extent alleviate or even solve the "information overload" problem.As the core of the recommendation system,the recommendation algorithm has a fundamental influence on the merits of the recommendation effect.With the development of the network,various enumeration recommendation algorithms are proposed and achieved remarkable results,but there are still some problems,especially in big data.In the background,the problems of low recommendation efficiency,poor scalability,and low recommendation quality are gradually emerging in the recommendation system,which needs to be solved urgently.In this paper,a Fusion algorithm based on clustering and matrix Approximation is proposed for the above problems.The specific work is as follows:(1)A co-clustering algorithm based on Bregman distance is used to mine low rank scoring sub-matrices with similar characteristics.By combining different constraint sets,different clustering calculation methods,and the number of rows and clusters,the multi-level and different levels of scoring sub-matrices are mined,paving the way for subsequent multi-model fusion,and the co-clustering algorithm splits the original scoring matrix.Improve the concurrency and scalability of the entire recommendation algorithm,and alleviate the high overhead of the algorithm.(2)The matrix Approximation is performed concurrently on the sub-matrix of each model.The SVD++ algorithm is used in the matrix Approximation stage.The weighting strategy is calculated based on the score distribution in each sub-matrix,and the high frequency score is given a larger weight.By sacrificing a small part of the data,the model is made to The high-frequency score of the scoring matrix is tilted to alleviate the under-fitting problems caused by insufficient sub-model data,and the prediction accuracy of most scores is improved,so that the recommendation performance is improved to some extent.The learning rate function is introduced in the gradient descent phase to control the update of the learning rate,thereby improving the efficiency of the algorithm.(3)The clustering algorithm determines the score distribution in the cluster,and the heterogeneous clustering produces different score distributions in the sub-matrix,that is,different sub-models are trained based on different clustering results.In this paper,the sub-model mean fusion method is used to multi-model.The fusion results into a single model to output the recommendation results,so that the recommendation results based on different clusters achieve complementary effects.Experiments show that the proposed algorithm has a significant reduction in root mean square error(RMSE)and mean absolute error(MAE)compared with the three baseline algorithms,that is,the recommended quality is greatly improved,and the recommended efficiency is also Corresponding improvement.