A Collaborative Filtering Recommendation Algorithm Based On Co-Pairwise Ranking

With the rapid development of the Internet and information technology,a large amount of information quickly floods into the Internet.While rich information brings convenience to users,it also leads to information overload.The field of information retrieval has always been considered as one of the effective ways to solve the problem of information overload,helping users to quickly obtain valuable information from massive information.Compared with the traditional information retrieval technology,the recommendation system can actively provide users with information that may be of interest without the clear needs of users.It has become one of the important tools to alleviate the problem of information overload,and has been widely studied and applied in the industry.Many e-commerce websites and multimedia platforms can easily embed personalized recommendation technology based on the existing system.For example,Amazon shopping website can help users find products of interest.Today's headlines can recommend news that may be of interest to the day.To a certain extent,the use of recommended technology not only increases the user's participation and trust and dependence on the application,but also brings considerable income to the merchants who use the technology,such as movies,news and POI recommendations.In real life,users' consumption behavior is complex and diverse,usually affected by many factors.The purchase decision made by the user is not only based on his/her own preferences,but also on the functional relationships between the historical purchase item and the item to be purchased.The main contributions of this paper are as follows:(1)Most of the existing recommendation algorithms only construct finer-grained models from the user's perspectives.Nevertheless,they overlook the functional complementary relationships among items--A key factor that can significantly influence user purchase decision-making process.Aiming at this problem,this paper construct sample pairs for specific users and items respectively from both user's and item's perspectives.We according to the combination of user-item interactions and item-item complementarity relationships,and propose a co-pairwise ranking model.(2)For the pairwise ranking method,the negative sampler will directly impact on convergence speed and the accuracy of recommendation.In order to speed up the convergence of the model,a novel rank-aware sampling strategy is constructed when constructing sample pairs for users and items according to the above two relationships.According to the ranking position of positive samples,the strategy selects more effective negative samples and defines the weight function learning gradient of dynamic control model(3)We devise an efficient collaborative filtering recommendation algorithm,named CPR(Co-Pairwise Ranking).The experimental results on four datasets show that CPR outperforms a series of state-of-the-art recommendation algorithms in many metrics(Precision,Recall,MAP and NDCG)and convergence speed.