Link Prediction Analysis Of Social Networks Based On Massive Data

A node in a network may represent a physical entity, and the edge connected could be mapped to the interaction between physical entities. In nature and human society, there exist lots of large and complex systems which can be described by various kinds of networks. If we map people in a social network to nodes and the relation between to edges, then a typical social network can be represented as a social graph G=<V,E>. By analyzing the social network topology and the node properties, link prediction tries to predict whether a specified link will occur by a given time in the social network.The introduction of supervised learning methods is the trend in link prediction. We can treat link prediction as a simple classification problem By obtaining link relationships of the current network and extracting feature sets of the links, we can take advantage of supervised learning models to classify those unknown links.Researching sampling methods to eliminate the impact of extreme class skewness, searching for algorithms of low computational complexity, designing models which can reflect the dynamic characteristics of social networks are three hot research areas in link prediction.Major works of this paper are as follows:1) The high complexity of massive data of social network analysis has become an unavoidable problem in link prediction, finding fast and efficient algorithms for similarity characteristics becomes the key to solve the problem above. This paper proposes a neighbor affinity measure, the experiment proves that that new measure improves the prediction accuracy, and also maintains a low computational complexity.2) Though1:1sampling methods successfully avoid the "drowned" problem caused by complex network sparsity, actual performance of the classifier trained is unstable, we need to find a new sampling method. Then we propose the13sampling strategy. The experiment shows that new sampling method effectively ameliorates the inadequate training issue.