Personalized Recommendation Based On Learning To Rank

How to carry out personalized recommendation efficiently and accurately from a large number of user data has always been an important research area.Sorting learning is a direction in this field.The main method is to train a large number of sample data through machine learning to generate a sorting function based on sample characteristics.In the field of learning to rank,there are three methods: pointwise,pairwise and listwise.In the process of learning to rank,the sample partial order pair often determines positive and negative cases by dichotomy,so as to distinguish the position relationship of samples.However,in the process of dichotomy,some sample data can not be distinguished well,so some sorting is difficult to determine.Aiming at the above problems,this paper puts forward an optimization scheme of sample pair relationship determination in pairwise.This scheme combines the theory of decision tree,introduces the score function to deal with the unclassified samples,so as to improve the accuracy of personalized recommendation.The experimental results show that the accuracy of SVM classification is improved with different penalty factors.On this basis,combined with boosting method,this paper proposes a sorting function optimization method DBSVR,which improves the accuracy of sorting function by setting the weight parameters of samples and classifiers.Aiming at the listwise,this paper proposes a new objective ListBN optimization function based on the scoring standard NDCG,which ensures the accuracy of the TOPN recommendation by giving more scores.For the data set MovieLens1 M,the DBSVR method proposed in this paper has advantages over both MPRank and SVRank,and the accuracy of TOP5 recommendation is increased by 2.98% and 16.1% respectively.Compared with ListNet and RankingSVM,the ListBN method proposed in this paper has higher accuracy of TOPN recommendation,and the accuracy of TOP9 recommendation is increased by 6.91% and 7.77% respectively.