An Adaptive Classification Framework for Data Streaming Anomaly Detection

Predicting the behavior of a system, we usually analyze its past data to discover common patterns and other classification artifacts. This process consumes considerable computational power and data storage. We propose an approach and a system, which requires much less resources without compromising prediction capabilities and accuracy. It employs three basic methods: common behavior graph, contour surrounding the graph, and entropy calculation methods. When the system is about to be implemented for a specific domain, the optimized combination of these three methods is considered, such that it fits the unique nature of the domain and its corresponding data type. In this work, we propose a framework and a process assisting system designers, finding the optimal methods for the case at hand. We demonstrate our approach with a case study of meteorological data collected over 15 years to classify and detect anomalies in new data.