Adaptive Anomaly Detection using Isolation Forest

摘要

This project developed an adaptive anomaly detection system based on Isolation Forest, applicable to data stream which demands single-scan online algorithms with poly-logarithmic time and space complexities. The proposed system based on Half-Space Tree, an extension of Isolation Forest, is not only capable of detecting anomalies when the underlying concept changes gradually over time, but also capable of detecting abrupt changes in the underlying concepts. Half-Space Trees is significantly better than three existing state-of- the-art distance-based and density-based methods, in terms of detection accuracy, time complexity and memory requirement.