An Ensemble Approach for Event Detection and Characterization in Dynamic Graphs

Event detection in datasets represented by dynamic graphs is an important task for its applications in a variety of domains, such as cyber security, online and telecommunications, fault and fraud detection, etc. Despite recent advances in this area, there does not exist a single winning algorithm known to work well across different datasets. In fact, designing a single method that is effective on a wide range of datasets is a challenging task. In this work, we propose an ensemble approach for event detection and characterization of dynamic graphs. Our ensemble leverages three different base detection techniques, the results of which are systematically combined to get a final outcome. What is more, we characterize the events; by identifying the specific entities, i.e. nodes and edges, that are most responsible for the detected changes. Our ensemble employs a robust rank aggregation strategy to order both the time points as well as the entities by the magnitude of their anomalousness, which as a result yields a superior ranking compared to the base techniques, thanks to its voting mechanism. Experiments performed on both simulated (network traffic flow data with ground truth) and real data (New York Times news corpus) show that our proposed ensemble successfully identifies the important change points in which a given dynamic graph goes through notable state changes, and reveals the key entities that instantiate these changes.