Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Jordan Levy, Paul Saves, Moncef Garouani, Nicolas Verstaevel, Benoit Gaudou

TL;DR

Unsupervised anomaly detection lacks labeled data, making robust ensemble design challenging. The paper proposes a SHAP-based framework to characterize detectors via input-feature attributions and to measure explainability-driven similarity, then links explanation diversity to ensemble performance using Mantel tests. It shows that explanation divergence signals complementarity and that explanation-based diversity can improve AUCPR when combined with strong base detectors, though diversity alone is not a panacea. The findings provide practical guidance for constructing more robust unsupervised anomaly detection systems and suggest extending the approach to time-series data and performance-informed selection.

Abstract

Unsupervised anomaly detection is a challenging problem due to the diversity of data distributions and the lack of labels. Ensemble methods are often adopted to mitigate these challenges by combining multiple detectors, which can reduce individual biases and increase robustness. Yet building an ensemble that is genuinely complementary remains challenging, since many detectors rely on similar decision cues and end up producing redundant anomaly scores. As a result, the potential of ensemble learning is often limited by the difficulty of identifying models that truly capture different types of irregularities. To address this, we propose a methodology for characterizing anomaly detectors through their decision mechanisms. Using SHapley Additive exPlanations, we quantify how each model attributes importance to input features, and we use these attribution profiles to measure similarity between detectors. We show that detectors with similar explanations tend to produce correlated anomaly scores and identify largely overlapping anomalies. Conversely, explanation divergence reliably indicates complementary detection behavior. Our results demonstrate that explanation-driven metrics offer a different criterion than raw outputs for selecting models in an ensemble. However, we also demonstrate that diversity alone is insufficient; high individual model performance remains a prerequisite for effective ensembles. By explicitly targeting explanation diversity while maintaining model quality, we are able to construct ensembles that are more diverse, more complementary, and ultimately more effective for unsupervised anomaly detection.

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

TL;DR

Unsupervised anomaly detection lacks labeled data, making robust ensemble design challenging. The paper proposes a SHAP-based framework to characterize detectors via input-feature attributions and to measure explainability-driven similarity, then links explanation diversity to ensemble performance using Mantel tests. It shows that explanation divergence signals complementarity and that explanation-based diversity can improve AUCPR when combined with strong base detectors, though diversity alone is not a panacea. The findings provide practical guidance for constructing more robust unsupervised anomaly detection systems and suggest extending the approach to time-series data and performance-informed selection.

Abstract

Unsupervised anomaly detection is a challenging problem due to the diversity of data distributions and the lack of labels. Ensemble methods are often adopted to mitigate these challenges by combining multiple detectors, which can reduce individual biases and increase robustness. Yet building an ensemble that is genuinely complementary remains challenging, since many detectors rely on similar decision cues and end up producing redundant anomaly scores. As a result, the potential of ensemble learning is often limited by the difficulty of identifying models that truly capture different types of irregularities. To address this, we propose a methodology for characterizing anomaly detectors through their decision mechanisms. Using SHapley Additive exPlanations, we quantify how each model attributes importance to input features, and we use these attribution profiles to measure similarity between detectors. We show that detectors with similar explanations tend to produce correlated anomaly scores and identify largely overlapping anomalies. Conversely, explanation divergence reliably indicates complementary detection behavior. Our results demonstrate that explanation-driven metrics offer a different criterion than raw outputs for selecting models in an ensemble. However, we also demonstrate that diversity alone is insufficient; high individual model performance remains a prerequisite for effective ensembles. By explicitly targeting explanation diversity while maintaining model quality, we are able to construct ensembles that are more diverse, more complementary, and ultimately more effective for unsupervised anomaly detection.
Paper Structure (21 sections, 5 equations, 2 figures, 7 tables)

This paper contains 21 sections, 5 equations, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised Anomaly Detection Algorithms
  4. Model Selection for Unsupervised Anomaly Detection
  5. Explainability
  6. Methodology
  7. Problem Setting
  8. Model Similarity from Explanations
  9. Linking Similarities to Detection Performances
  10. Experiments
  11. Experimental Setup
  12. Datasets Considered
  13. Correlation Between Similarity Matrices
  14. Aggregate Models Predictions
  15. Complementarity
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Mean similarity between models across all datasets. Layout: Top-Left: correlations of SHAP values; Top-Right: NDCG of SHAP values; Bottom-Left: correlations of anomaly scores; and Bottom-Right: Jaccard similarities.
  • Figure 2: Relationship between ensemble diversity (given by $\rho^{PS}$) and average individual performance. Each point represents an ensemble of models. The color scale indicates the ensemble's overall performance (AUCPR).