| With the emergence of various Internet applications, information over-load problem has become more and more serious, and recommendation system has gradually become one of the most important technologies to address this problem. Although recommendation system has been studied for many years, and been successful in many fields, but most of recom-mendation systems still suffer from their inherent defects and deficiencies,such as: cold start problem, data sparse problem, and lack of making full use of the context data, etc. To overcome the inherent problems of traditional recommendation system, academia and industry are trying to integrate social information, which is collected by online social networks,into traditional recommendation system so as to enhance and improve its performance. Social recommendation is a new field, and this research fo-cuses on the social recommendation problem in the sense of generalized definition, and chooses four research points to carry out the innovative research works:Firstly, the measurement of similarities between users is an important issue in social recommendation. Aiming at the characteristics of large scale and sparseness of data in social recommendation, this paper proposes a new social recommendation method based on manifold ranking and ma-trix factorization, which integrates the similarity between users into the low-order matrix factorization of rating matrix. Experiments show that on the premise that the manifold structure of users' implicit features is consistent with the social network structure, the proposed method has higher recommendation accuracy and lower RMSE and MAE values than the current similar methods on a large data set.Secondly, the majority of existing researches on social recommenda-tion emphasize the commonness, while ignoring the non-social impact on user behavior. In this research, a new social recommendation method based on matrix factorization is proposed, in which social relations, rating historical data are integrated into the objective function with appending additional penalty terms and bias terms to classic matrix factorization model. The experimental results show that under the condition that most users and items manifest both sociality and non sociality, the proposed method has better performance than other similar methods.Thirdly, most of friend recommendation methods make recommen-dation by exploiting social relations and interest similarities implied in the social relations, while few of them explicitly incorporate ratings into recommendation in addition to exploiting social relations. Therefore, this paper studies the friend recommendation problem considering both so-cial relations and similar interests, and proposes a matrix factorization based friend recommendation method. This method takes into account both social relations and ratings, where Gaussian kernel is used to capture the interest correlation between users and the user's latent feature vector is generated by Gaussian process, thus forms a friend recommendation method which has the capability of recommending friends with common interests in social network. Experiments show that, with the precondi-tion that social networks have implicated few information of interest char-acteristics, our method outperforms the art-of-the-state traditional link prediction methods using only social relations.Finally, analysis of social tagging is an important way to understand user preferences, and is even better than the analysis of user ratings and social relations. Therefore, this paper studies the social recommendation based on tag information, rating records and social relations, and proposes a social recommendation method with tag side information. This method decomposes the rating matrix, tagging matrix, and annotating matrix si-multaneously by means of matrix factorization and regularization method,with social constraints being imposed on latent feature of users. Experi-mental results have justified the effectiveness of the proposed algorithm. |
PDF Full Text Request