Rule Extraction in Unsupervised Anomaly Detection for Model Explainability: Application to OneClass SVM

Unsupervised anomaly detection models lack transparency: OneClass SVM, a widely-used method for detecting unusual patterns in data, operates as a "black box"—meaning organizations cannot easily explain why the system flagged something as anomalous, which is a significant barrier to responsible deployment in high-stakes applications.

Rule extraction can make anomaly detection interpretable: The paper demonstrates that extracting simple, human-readable rules from OneClass SVM models is feasible, offering a practical pathway to explain anomaly detection decisions to stakeholders and end-users without requiring complete model redesign.

Explainability for unsupervised learning is underdeveloped: While explainability techniques have advanced significantly for supervised models (like classifiers), the paper identifies that unsupervised learning—which is increasingly used for anomaly detection—remains largely unexplored in XAI research, representing a gap in responsible AI practices.

Multiple dimensions of explainability matter: The authors propose measuring rule quality across four dimensions—comprehensibility (ease of understanding), representativeness (coverage of decisions), stability (consistency over time), and diversity (variety of explanations)—suggesting that good explanations require balancing multiple factors rather than optimizing a single metric.

Real-world validation strengthens practical applicability: By testing their approach on industrial datasets alongside academic benchmarks, the authors demonstrate that rule extraction techniques can work in actual business contexts, making this a step toward deployable explainability solutions for anomaly detection systems.