Research On Multi-features Link Prediction Based On Matrix

Data mining is a kind of technique that combines data analysis and complex algorithms for processing huge data sets. Currently, as a branch of data mining, link mining's research tasks include description and prediction models, and with the introduction of link concept, link prediction is now becoming hotspots of this area.Nowadays, on the one hand existing link prediction methods have only used a single feature (structural feature or attributive feature) for link prediction, and ignored the time factor and the importance of features; on the other hand, they cannot systematically extract topological features and semantic features for given networks, so they get worse performance. In allusions to these problems, the paper does some research as followed.Firstly, in order to simultaneously consider attribute information, structural information and network dynamics provided by social networks, this paper uses matrix methods to analyze various kinds of information from networks. And we give the representation of social network and its information.Secondly, this paper presents a matrix-based approach about link prediction, which combines temporal feature, weighted attributive features and weighted topological features. This approach uses an alignment relation of social network to identify important features, and then effectively combines these features in a matrix-based method. This approach can improve the performance of link prediction. In addition, this paper uses the singular value decomposition to reduce the storage space.Thirdly, for co-authorship networks, we first give the definition of link prediction problem in co-authorship networks, and for capturing effective information about networks from many aspects, we extracted topological features, semantic features and temporal features systematically and step by step from the given networks, and then presented a link prediction model which simultaneously utilized these three features by combining these features using supervised learning framework. It can improve the predictive performance of link prediction again.Finally, the experimental results confirme the validity and feasibility of the algorithms and the anticipated results are realized.