Towards Explainable Anomaly Detection: Optimal Counterfactual Explanations in Isolation Forests

Isolation Forest (IF) is a widely used and efficient method for unsupervised anomaly detection, but its internal decision process is not easy to interpret, partly due to the random feature and threshold choices in the trees. This lack of clarity can make it difficult to understand why a specific point is marked as anomalous. To address this challenge, this thesis proposes an approach that adapts the Optimal Counterfactual Explanations (OCEAN) framework to the unsupervised nature of IF. OCEAN, originally developed for explaining classifications in tree-based ensembles, is repurposed here to generate counterfactual explanations for anomalies. In practice, the method finds how an anomalous data point’s features could be minimally modified so that the IF model would consider it normal. The proposed counterfactual approach is evaluated against DIFFI, a recent model-specific interpretability technique for IF that uses depth-based feature importance scores. Experimental results indicate that in several cases OCEAN-based counterfactuals and DIFFI attributions agree on the influential features driving an anomaly, while in other cases the explanations diverge and offer different insights. In summary, this thesis extends explainable AI to unsupervised anomaly detection by introducing a counterfactual interpretability method for Isolation Forest, with the aim of improving transparency and user trust in anomaly detection models.