Algorithm Research Based On The Cold Start Problem Of Recommender System

With the rapid development of the Internet,it is more and more difficult for people to quickly locate the information they need in all kinds of information oceans.In order to help users find the information they need from the massive information more quickly and conveniently,the recommendation system has become the key Part of it.Personalized recommendation for users has always been a hot topic in the recommendation system research,but when faced with a user who has just entered the system,the system does not have any data about the user,that is,a "cold start user",and the recommendation system cannot be personalized at this time recommend.Using cross-domain recommendation to alleviate the cold start problem is a current research hotspot.This method provides a new research direction for solving cold start.Cross-domain recommendation is to use data from different domains to learn the relationship between domains to make recommendations for cold-start users.This article is mainly aimed at the user's cold start problem,researching on the basis of cross-domain recommendation,and proposing a cross-domain recommendation model based on matrix three factorization.The main content of this article is the following four points:(1)This article first reviews and summarizes the significance of the cold start problem in the research recommendation system and the research process of scholars on the recommendation system and the cold start problem,introduces four classic recommendation algorithms,and then elaborates the theoretical knowledge of the cold start problem,To introduce algorithms to alleviate the cold start problem:non-personalized recommendation,social network recommendation,tag recommendation.Although these methods have made a great contribution to alleviating the cold start problem,there are still big problems.Just like in cross-domain recommendation,when decomposing the potential features of the domain,only the binary relationship between the user and the item is considered,and the characteristics of the domain itself are ignored.At the same time,in the data set,there are users who have nothing to do with cold start users in the scoring matrices of the two domains,and these users will bring noise to the decomposition results.(2)In response to the above problems,this paper proposes a cross-domain recommendation model based on matrix three factorization,which is mainly aimed at improving the calculation model of potential characteristics in cross-domain recommendation.The traditional latent characteristic calculation model uses the matrix decomposition method to establish the user-item binary relationship.In this paper,the domain factor is added in the matrix decomposition to construct the user-domain-item ternary relationship.Through matrix three factorization,the domain potential feature matrix is decomposed,and the relationship between domain factors and users and items is captured.(3)In the past cross-domain recommendations,for commonly used data sets that contain users unrelated to cold start users,this paper proposes to filter the data set first,set thresholds,and eliminate irrelevant users to optimize the data set.At the same time,the dimensionality of the filtered data set is reduced,which is more conducive to experimental operations.In addition,in the latent feature learning module,this paper uses multi-layer perceptrons and gradient boosting trees as learning models to implement CDMTF-MLP and CDMTF-GBDT models.(4)This paper conducts experimental operations on two real data sets of MovieLensNetflix and Amazon to verify the feasibility of the cross-domain recommendation model based on matrix three factorization,and compare and analyze it with important models in cross-domain recommendation.The experiment sets RMSE and MAE as evaluation indicators,and compares the RMSE and MAE values on the two data sets from different angles,which proves that the improved method of the model in this paper is reasonable,effective and feasible compared with the previous traditional models.In addition,for data preprocessing,experiments were conducted on the impact of overlapping users on the accuracy of the target domain cold-start user rating prediction,and the experimental results were visually displayed and analyzed.