Reproduce “Interpretable Anomaly Detection with DIFFI: Depth-based Isolation Forest Feature Importance” on OpenResearch

About this paper

Mattia Carletti, Matteo Terzi, Gian Antonio Susto

Researchers at the University of Padua developed DIFFI (Depth-based Isolation Forest Feature Importance), a model-specific framework that provides global and local explanations for Isolation Forest anomaly detection. This method demonstrates significantly faster computation for local explanations compared to SHAP, and achieves competitive or superior performance for unsupervised feature selection across various datasets.

Key insights

• Feature importance in Isolation Forest is directly linked to a feature's ability to isolate anomalous data points quickly at shallow tree depths, and to create highly imbalanced splits.
• Model-specific interpretability methods, tailored to an algorithm's internal structure, can provide more faithful and accurate explanations than general model-agnostic approaches.
• Developing a post-hoc interpretability method allows for preserving the high predictive performance of established anomaly detection algorithms while offering necessary transparency.

Summary by alphaXiv.