Research And Implementation Of Recommendation Algorithm Based On Determinant Point Process

A good recommendation prediction can reduce the user’s search space and help the user’s decision-making process.In recent years,accurate matching of user preferences has always been the direct driving force in the field of recommendation,but the accuracy of recommendation alone is not enough to measure the actual effectiveness of the recommendation effect and the user’s consumption experience.Repeated pushes of the same type may limit users’ vision and lead to a decline in user satisfaction.More and more scholars have reached a consensus that recommendation diversity should be included as a key dimension in evaluating recommendation utility,to better enhance the user experience with the recommendation system.For the Top-N recommendation task,this paper mainly completes the following three aspects.First,a recommendation diversity algorithm FIBE based on the Determinantal point process(DPP)is proposed.Based on the global negative correlation of the repulsive probability model,it greatly improves the diversity of the recommendation list and ensures recommendation relevance.The kernel matrix constructs of two parts,the set diversity measured by the pairwise personalized ranking loss,and the candidate quality based on the aggregated similarity.The algorithm combines the backward greedy method in the DPP maximum a posteriori inference to help the method get rid of the local optimum.Second,an efficient determinant point process acceleration algorithm AFIBE based on Givens rotation and Cholesky transform is proposed.The recommendation model based on DPP needs a large number of determinant calculations,and the application of backward greedy in the process of maximum a posteriori inference involves the elimination of specific rows and columns of the determinant index by the candidate element,and hence the determinant point process based model is time consuming.This paper leverages Givens rotation and Cholesky decomposition in the DPP maximum a posteriori inference,to transform the principal minor of the kernel matrix indexed by the elements of the recommended subset,which enables it to be solved and updated incrementally with the expansion of the recommended subset.At the same time,according to the submodular properties of DPP maximum a posteriori inference,the lazy calculation is used to reduce the amount of iterative calculation to meet the real-time requirements of the recommendation system.Third,this paper designs and builds a recommendation prototype system.The recommendation system uses Flask and Streamlit as the architecture to build web applications and uses MySQL and Redis databases for hierarchical storage.In this paper,the proposed recommendation diversity enhancement algorithm based on DPP is fully verified,and the conclusion indicates that FIBE ensures a substantial improvement over the accuracy-diversity tradeoff,and the performance is improved compared with the comparison algorithm.Meanwhile,the computational efficiency of the accelerated algorithm AFIBE is also better than that of the benchmark methods.