Skip to content
Book Open access

XGIF: An Explanation-Guided Anomaly Detection Extention of Isolation Forest

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 15 references

Abstract

Isolation Forest ranks among the most widely used unsupervised anomaly detection algorithms due to its simplicity, scalability, and computational efficiency. However, its tree construction process relies exclusively on randomly selected splitting features and split values, ignoring informative signals about how anomalies are characterized in the feature space. In this paper, we propose XGIF, an explanation-guided extension of Isolation Forest that exploits feature attribution signals to drive the isolation process. Our core insight is that anomalous instances tend to exhibit unstable attribution patterns across the tree ensemble, while normal instances are characterized by more consistent attributions. XGIF introduces attribution stability as a novel criterion for evaluating leaf purity and adaptively extends tree growth when unstable explanatory patterns are detected. Furthermore, the growth process is guided toward features that contribute most to anomaly characterization, enabling more effective isolation of outlying instances. By integrating explainability directly into the learning mechanism, rather than confining it to post-hoc interpretation, XGIF leverages attribution signals as an active driver of anomaly detection performance. Experiments on 20 benchmark datasets from ADBench demonstrate that XGIF significantly improves detection performance, outperforming Isolation Forest on 17 of 20 datasets.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.