Information Mining Research On Multi-dimensional Social Networks

With the thriving of the Internet, social networks have already become an essential part of our daily life. And the multiple social networks existing at the same time tend to merge as one united multi-dimensional social network. Information mining research on multi-dimensional social networks is valuable and very meaningful for academic literature and practical applications. However, traditional information mining algorithms and models that applied to single social network are hardly to be employed in the scenario of multi-dimensional social networks. This thesis will focus information mining problems on multi-dimensional social networks.First, transferring-similarity-based recommendation on multi-dimensional social networks. Information recommendation on multi-dimensional social networks is facing the data missing problems, i.e. data sparsity and cold-start. In the meantime, there are rich information on each dimension of a multi-dimensional social network, however, methods that can be employed to conquer the data missing problem and to enhance recommendation performance are still undiscovered. In this thesis, a transferring-similarity-based recommendation algorithm designed for multi-dimensional social networks is proposed. This algorithm improves the recommendation performance and solves data sparsity and cold-start problems on the target social network by leveraging the information from other social network dimensions, through the bridge composed by shared uses between different social network dimensions. An accelerating algorithm of similarity computation is also proposed. It reduces the time complexity significantly by a small space complexity cost. It makes the proposed recommendation algorithm more practical. The effectiveness of the proposed recommendation algorithm is proved by experiments performed on a data set of multi-dimensional social network of real e-commerce websites. The proposed algorithm outperforms traditional user-based collaborative filtering algorithm with a significant accuracy improving(1-2 times).Second, information propagation prediction on multi-dimensional social networks. In this thesis, influence factors from three dimensions, i.e. social connection network, message passing network, temporal feature are taken into consideration. A temporal-feature-based probability prediction model of information propagation is proposed. This model does not only involve the static factor of topologic structure of connection network, but also creatively involve the temporal feature of message passing network. In the inference process of this model, prior probability distributions are assigned to model parameters, then maximum a posteriori estimate strategy is employed to learn model parameters. In order to examine the information propagation prediction performance of this proposed model, experiments are conducted on a Weibo data set. Results prove that this proposed model outperforms other mainstream prediction models with significant higher accuracy, even with fewer training samples, this proposed model can still provide predictions with higher accuracy.