Modeling entities in interaction dataset for anomaly detection and explanation

Interaction datasets generated from online social networking, network communications, phone calls or emails contain enormous information about entities and their behaviors. Detecting anomalies, whose behaviors derivate a lot from other majority, plays significant role in artificial systems. In this work, we propose a new method to model entities in such context, capturing dynamic behaviors elastically. The modeling algorithm, called SPB (Soft Plastic Ball) is detailed, which is scalable and sensitive to variations of each entity behaviors, so as to detect anomalies timely. A distance-based outlier detection algorithm is applied to uncover anomalous entities. A visualized explanation of results is also provided to alleviate false-positive problem, and aids people to gain more actionable information. Validity and effectiveness of our proposal is proved by experimental results.