The Research And Applications On Cross-Domain Recommendation System Based On Social Network

Information recommendation technology, which is based on the theory of data mining and pattern recognition theory, is able to predict information from given data. Meanwhile, in accordance with the different interests of users, the recommendation system, by the detailed analysis of specific issues, can provide differentiated and personalized service. From the actual historical data of users, the recommendation system can analyze the natural patterns within and forecast the unknown tendency; this is the normal principles and solutions of the recommendation technology. Users do not need to make a clear demand to the recommendation system, because it can initiatively extract the requests or preferences from the large data set. The technologies of Cross-domain recommendation system are the main related theories and procedures, which put the information fusion into practice, and combine the advantages and features from data sets within various domains.Although a variety of recommendation system can be found from various Internet services, the effectiveness of them cannot be fully satisfied. In other words, the recommendation system is facing manifold challenges from many aspects currently. To settle these challenges, this paper has studied the prediction methods of social network link. Combined with the existing theory and technology, we have applied them to the pragmatic recommendation system. Specifically, this project has mainly completed the following work. (1) Before the startup of recommendation systems, in order to solve the challenges of the data sparsity problem, we have done the researches on how to reduce the sparsity of data by the exploit of data on social network. Also, the network link prediction methods are applied to calculate the similarity of each node, i.e. by using information fusion approaches, the performances of recommendation could be improved.(2) The methods on the similarity of neighboring sub-network have been proposed, as well as the methods on the similarity based on the degree ratio of the sub-network to the whole network. And the methods have been applied to compute the similarity of social network users. By adopting these methods, the accuracy and the outcome of node similarity based on link prediction methods have been advanced.(3) The algorithms of similarity fusion have also been studied. And some new algorithms have been proposed to improve the accuracy of user similarity judgment. These new methods are based on the sparsity of data of both the social network and the original user-item ratings.(4) The Latent Factor Model which is based on Singular Value Decomposition (a matrix factorization method) are studied and applied to build the recommendation systems. In order to avoid over-fitting phenomenon which occurs when training the models, the objective function has been optimized, by adding the weight decays of the similarity of users. The link prediction algorithms are employed to regularize and overcome the defects of the fitting method.In this study, we have analyzed and optimized the imperfections and properties of the current recommendation algorithms in several aspects. Combined the new methods with existing theories and algorithms, we have also designed the testing methods.In order to evaluate the performances of algorithms or methods, on the basis of evaluating indicators, we have introduced controlled trials to verify their accuracy and validity. In this project, the applications of information fusion technology to solve problems and improve the existing methods can have important significance for the construction technology in the recommendation systems.