Research On Matrix Factorization Recommendation Algorithm Of Base On Feature Overcross

With the development of Internet science and technology and the advent of the information age,people's demand for information has also increased.However,the difficulty of finding information has not decreased with the advent of the information age,but has increased.As a technology that can help people find information about their own needs,recommendation systems have received extensive attention and research.Collaborative filtering,as the earliest recommendation algorithm,has gradually been unable to meet people's needs.People need recommendation algorithms with better performance.Provide referral services.Under this background and purpose,this article through in-depth exploration of the matrix factorization model and factorization machine model,summarizes and analyzes their advantages and disadvantages,combines the algorithm ideas of matrix factorization and factorization machine,and proposes a matrix factorization that integrates feature intersections.Recommended algorithm.The main innovation of the algorithm proposed in this thesis is to integrate the idea of feature cross combination into the prediction calculation of the matrix factorization model,and to avoid learning all cross terms between each feature dimension separately,all feature cross coefficients are merged into one feature cross term Coefficient matrix,by learning this matrix to capture the interactive relationship between various features.The algorithm is based on the matrix factorization model,replacing the calculated vector inner product in the matrix factorization model as a predictive score with the effect of calculating the cross combination between the features of the vector on the score result,and retaining the original matrix factorization framework using the gradient descent method The final hidden feature matrix and feature second-order cross term coefficient matrix are calculated to form a matrix decomposition recommendation algorithm for fusion feature cross.Experiments on the Movie Lens data set.In terms of predicting scoring errors,the algorithm in this thesis is compared with other matrix factorization recommendation algorithms.On the ml-1m dataset,the accuracy rate is increased by 2.21% compared with that of Funk factorization recommendation algorithms.The experimental results show that the algorithm in this thesis has better performance,indicating that the fusion feature crossover can bring a certain improvement to the matrix factorization algorithm.Conducive to better meet the needs of users for recommendation algorithms.