KNNIFE

Riccardo De Vidi, Francesco Borsatti, Valentina Zaccaria, Davide Sartor, Gian Antonio Susto

---

Code for the paper KNNIFE: k-Nearest Neighbors Isolation Forest Interpretability.

KNNIFE and sKNNIFE are feature importance methods for Isolation Forest-based anomaly detectors (IF, EIF, DIF). They assign importance scores to features by using k-nearest neighbor distances at each tree node split, and are benchmarked against DIFFI and ExIFFI.

Structure

• MyEnsemble.py — wrappers for IF, EIF, and DIF models
• MyInterpreter.py — GFI/LFI computation for DIFFI, ExIFFI, KNNIFE (kNN), and sKNNIFE (d)
• diffi_interpretability_module.py — DIFFI baseline implementation
• my_datasets/ — benchmark datasets
• performance/performance.ipynb — anomaly detection performance (ROC-AUC, AP, etc.) for IF/EIF/DIF
• interpretability_evaluation/GFI_evaluation.ipynb — Global Feature Importance and AUC-FS evaluation for all methods