Online Rare Events Detection

Rare events detection is regarded as an imbalanced classification problem, which attempts to detect the events with high impact but low probability. Rare events detection has many applications such as network intrusion detection and credit fraud detection. In this paper we propose a novel online algorithm for rare events detection. Different from traditional accuracy-oriented approaches, our approach employs a number of hypothesis tests to perform the cost/benefit analysis. Our approach can handle online data with unbounded data volume by setting up a proper moving-window size and a forgetting factor. A comprehensive theoretical proof of our algorithm is given. We also conduct the experiments that achieve significant improvements compared with the most relevant algorithms based on publicly available real-world datasets.