Research On Link Prediction Method Using Community Information

With the development of information society, more and more social networks with mass data appear. Mining the valuable information from these networks becomes a challenging task. Link prediction is not only a sub-area of data mining but also a key task of social network analysis. Link prediction aims to use the existing information in network to predict the relationship that could or will be established, which makes it significant in theory or in real application to understand the network structure and to find the hidden information in the network. Recently, link prediction has become a hot topic of data mining and is widely used in various real networks such as social networks, biological networks, information networks etc.Current, most of the link prediction methods are based on the node similarity due to its low time complexity and good performance of prediction. Afterwards, some researchers use supervised learning technique in link prediction to model the formation of links based on structual features.However, these methods have limitation in that they only consider local relationship of the current node withits neighbor nodes while ignore the global relationship of the current node with its non-neighbor nodes. In fact, real networks usually have community structures where the links among nodes are close in the same community but loose in different communities. Community has important impact on the formation of links and generally node pairs in the same community are easier to connect. Therefore, it deserves in depth research on how to use the community information in the networks to improve existing link prediction methods for a higher prediction accuracy.Considering the problems and challenges in link prediction, we propose to use community informationin link prediction task. Based on the supervised learning framework, we propose a new link prediction method using community information to reform instance features(CR-LiP) and also another new link prediction method using community information to expand instance features(CE-Li P). Finally, we conduct experimental simulation using ACF and FaceBook dataset to compare the CR-LiP and CE-LiP methods with some benchmark methods and other related methods.Experimental results show that the performance of a supervised link predictor can be improved when we use community information extracted from networks.