Analyzing Shapley Additive Explanations to Understand Anomaly Detection Algorithm Behaviors and Their Complementarity

Unsupervised anomaly detection remains a core challenge in machine learning, as the lack of labels and diverse data distributions make it hard to build robust detectors. Ensemble methods—combining multiple detectors—aim to reduce individual biases and improve reliability, but often fail because many models rely on similar decision cues, producing redundant anomaly scores. A new paper from researchers at Université Toulouse Capitole and CNRS tackles this by proposing a methodology to characterize detectors through their decision mechanisms using SHAP (SHapley Additive exPlanations).

By quantifying how each model attributes importance to input features, the team creates attribution profiles that measure similarity between detectors. Their key finding: detectors with similar explanations produce correlated scores and identify overlapping anomalies, while explanation divergence reliably indicates complementary detection behavior. This explanation-driven metric offers a different criterion than raw outputs for selecting ensemble members. The study also confirms that diversity alone is insufficient—high individual model performance remains a prerequisite for effective ensembles. By explicitly targeting explanation diversity while maintaining model quality, the researchers constructed ensembles that were more diverse, complementary, and ultimately more effective for unsupervised anomaly detection. The work earned the IDA 2026 Frontier Prize and Best Paper Award from Springer Nature.