Anomaly Subgraph Mining in Large-Scale Social Networks

Anomaly detection is a relevant research field in online social networks (OSNs), which detects the abnormal individuals and activities to protect other regular activities of the whole network. Current detection approaches can be divided into two types: structure-based automatic detection and experience-based manual detection. However, the first approach may have inaccurate detection results and neglect some abnormal nodes; The second one needs a great deal of experts' experience and needs such high labor costs. In this paper, we design a training-based graph pattern matching systems (T-GPM) to detect abnormal individuals in large-scale social networks. T-GPM first mines the feature patterns by training some existed abnormal samples. Then, by applying the feature patterns to the original social networks, T-GPM can find out all the similar abnormal nodes or subgraphs. Our experiments show that T-GPM can get essential features of the abnormal sample without extra experts' experience. Meanwhile, the accuracy of T-GPM increases with the number of iterations, and it achieves the final accuracy rate of more than 90%. Finally, the entire system has only a small amount of labor cost for subgraph filtering in the predicting phase.