Research On Social Recommendation Algorithm Based On Implicit Similarity And Metric Learning

Traditional recommendation systems usually focus on making full use of user-item rating information.With the continuous expansion of the application scenarios of the recommendation system,it is often impossible to obtain satisfactory recommendation results by making relevant recommendations based only on the binary rating relationship between users and items.The social recommendation system uses social information as auxiliary data,and attempts to improve the accuracy of the recommendation system's rating prediction and item ranking recommendation tasks by considering the social relationship and rating data in the social network.At present,scholars have achieved a series of research results on social recommendation algorithms,but there are still the following deficiencies:(1)Although the social relationship has been integrated into the calculation of similarity,the implicit information in the social relationship and rating data has not been fully tapped;(2)The social recommendation algorithm uses the dot product method in matrix factorization.The inherent shortcomings of this method will cause errors in the recommendation results;(3)Rating prediction and item ranking are two major tasks of the recommendation system.Although the accuracy of recommendation can be improved by integrating into social relationships,the basis for ranking items is still the rating information.In response to the above problems,this paper focuses on the following work:(1)A recommendation algorithm fusing implicit similarity of users and trust is proposed in this paper.The algorithm also considers the similarity between a user and his friends when the user as a truster and a trustee,and indirectly implements trust transfer.Therefore,the implicit information of rating information and trust relationship is more fully utilized.At the same time,the algorithm uses the user feature vector to calculate the user's rating similarity without considering the common rating set;the trust relationship feature vector is used to calculate the user's social relationship similarity without considering the common friend set,effectively eliminating the data sparsity problem caused by the lack of common objects in the traditional similarity function.In addition,the algorithm performs adaptive weighting constraints on the regularization term to prevent overfitting and minimizes the correction error.A series of comparative experimental results on the two data sets show that this method effectively improves the accuracy of the rating prediction of the recommendation system,and can effectively alleviate the problems of data sparsity and cold start.(2)A social recommendation algorithm based on metric learning is proposed in this paper.The algorithm can be used in two classic application scenarios: rating prediction and item ranking,which correspond to the social recommendation model for rating prediction and the social recommendation model for item ranking,respectively.The distance metric learning is introduced into the algorithm,which makes up for the deficiency of the matrix decomposition of dot product mode,and makes the whole process more intuitive and explanatory.The method implements the recommendation tasks of rating prediction and item ranking in four steps.First,convert the preference matrix(rating matrix or implicit feedback matrix)into a distance matrix;Secondly,a confidence mechanism is introduced to set a higher confidence level for relatively reliable ratings,and at the same time,the distance matrix is decomposed through metric learning to embed users and items into a unified low-dimensional space;Then,through the distance constraint,the distance between the user with the preferred items and trusted friends is shortened,and the distance between the user with the disliked items and untrusted friends is pushed away,thereby fully exploiting the influence of social relationships on the distance;Finally,after obtaining the positions of users and items in the low-dimensional space,it is applied to two application tasks of rating prediction and ranking recommendation,respectively.A series of comparative experimental results on the two data sets show that this method effectively improves the accuracy of rating prediction,improves the precision and recall of item ranking recommendations,and can also effectively alleviate the cold start problem.In summary,the problems of data sparseness,cold start,social relations,and insufficient mining of implicit information in social relations and rating data have been well solved in this paper.The proposed recommendation algorithm fusing implicit similarity of users and trust provides a good solution for solving the similarity calculation problem in the social recommendation system;The proposed social recommendation algorithm based on metric learning provides a new method for the study of how to better integrate social relationships with traditional recommendation systems,the rating prediction task and item ranking recommendation task in the social recommendation system have also been effectively improved and optimized.