Reproduce “Function Based Isolation Forest (FuBIF): A Unifying Framework for Interpretable Isolation-Based Anomaly Detection” on OpenResearch

About this paper

Alessio Arcudi, Alessandro Ferreri, Francesco Borsatti, Gian Antonio Susto

Researchers from the Universities of Padova and Sapienza introduce the Function-based Isolation Forest (FuBIF), a unified framework for isolation-based anomaly detection that incorporates arbitrary real-valued functions for data partitioning. This framework, along with its FuBIF Feature Importance (FuBIFFI) algorithm, provides competitive anomaly detection performance, often surpassing Deep-IF and AutoEncoder on average, and offers consistent, probabilistically grounded feature importance for enhanced interpretability across diverse datasets.

Key insights

• The optimal splitting function for isolation-based anomaly detection is dataset-dependent, highlighting the need for flexible frameworks that can adapt to varying data manifold structures.
• Generalizing Isolation Forest's splitting mechanism to arbitrary functions allows for unifying existing variants and systematically designing new ones, enhancing model adaptability.
• Probabilistically grounded feature importance, derived from split probabilities, provides consistent and actionable interpretability across diverse complex splitting functions, addressing a key XAI challenge.

Summary by alphaXiv.