Anomaly Detection with Generalized Isolation Forest

The existing Isolation Forest algorithm detects anomalies by using binary trees. However, a few limitations can be observed. To resolve this problem, we present a systematic approach to generalize and improve the Isolation Forest algorithm in this paper. The resulting Generalized Isolation Forest algorithm is based on the identification of common properties that characterize successful variations of Isolation Forest. It is designed to detect different kinds of anomalous data. Our analytical evaluation shows that the generalization conforms to the identified properties. Our experimental evaluation on real-world data demonstrates that our generalized method can significantly improve the performance of Isolation Forest when compared to the basic implementation. Overall, our work provides a foundation for understanding the fundamentals of Isolation Forest and introduces novel ways to extend the algorithm for improved performance in specific application domains.