Abstract
Numerous algorithms have been proposed for detecting anomalies (outliers, novelties) in an unsupervised manner. Unfortunately, it is not trivial, in general, to understand why a given sample (record) is labelled as an anomaly and thus diagnose its root causes. We propose the following reduced-dimensionality, surrogate model approach to explain detector decisions: approximate the detection model with another one that employs only a small subset of features. Subsequently, samples can be visualized in this low-dimensionality space for human understanding. To this end, we develop PROTEUS, an AutoML pipeline to produce the surrogate model, specifically designed for feature selection on imbalanced datasets. The PROTEUS surrogate model can not only explain the training data, but also the out-of-sample (unseen) data. In other words, PROTEUS produces predictive explanations by approximating the decision surface of an unsupervised detector. PROTEUS is designed to return an accurate estimate of out-of-sample predictive performance to serve as a metric of the quality of the approximation. Computational experiments confirm the efficacy of PROTEUS to produce predictive explanations for different families of detectors and to reliably estimate their predictive performance in unseen data. Unlike several ad-hoc feature importance methods, PROTEUS is robust to high-dimensional data.
Introduction
Unsupervised anomaly detection is a cornerstone of modern data science, used to identify everything from credit card fraud to sensor failures. However, most state-of-the-art detectors, such as Isolation Forest or Local Outlier Factor (LOF), are "black boxes"—they flag a record as anomalous without explaining the underlying cause. This lack of interpretability makes it difficult for human analysts to diagnose root causes or trust automated decisions.
To address this, we propose a surrogate model approach. The core idea is to approximate the complex decision boundary of an unsupervised detector using a simpler, supervised model that relies on only a small subset of features. This reduced-dimensionality space allows for intuitive visualization and global understanding of the detector's logic.
The Need for Predictive Explanations
Existing explanation methods typically fall into two categories: local explanations (explaining individual points) and descriptive explanations (explaining historical data). These approaches often fail to provide a high-level rule for the entire dataset or cannot be applied to new, incoming data without re-running expensive processes.
We introduce the concept of predictive explanations. A predictive explanation identifies the specific "subspace" (subset of features) that defines an anomaly and creates a lightweight model that can monitor future data streams in real-time. By treating explanation as an approximation problem, we can measure the quality of an explanation by its ability to predict future detector decisions.
The PROTEUS Pipeline
We developed PROTEUS, an AutoML pipeline designed to transform unsupervised detectors into interpretable, predictive models. PROTEUS is specifically engineered to handle the extreme class imbalance inherent in anomaly detection.
Key Components of PROTEUS:
- Labeling & Oversampling: An unsupervised detector (the "teacher") labels the dataset. To probe the decision boundary, PROTEUS generates "pseudo-samples" by adding Gaussian noise to known anomalies and asking the teacher to label them.
- AutoML Search: The system evaluates over 1,800 configurations of feature selection algorithms (e.g., LASSO, SES) and classifiers (e.g., SVMs, Random Forests) to find the smallest feature set (typically $\le 10$) that best mimics the detector.
- Performance Estimation: To avoid over-optimistic results, PROTEUS uses Bootstrap Bias Correction (BBC). This ensures that the reported accuracy (AUC) is a realistic estimate of how the explanation will perform on unseen data.
- Visualization: The pipeline produces Spider Anomaly Explanation (SAE) charts, which use polar plots to show how an anomaly's feature values deviate from the normal interquartile range.
Experimental Results
PROTEUS was evaluated using synthetic datasets with hidden "gold standard" features and real-world benchmarks (Wisconsin Breast Cancer, Ionosphere, and Arrhythmia) with injected noise.
- Feature Identification: On synthetic data, PROTEUS consistently achieved near-perfect recall of relevant features, significantly outperforming ad-hoc methods like SHAP.
- Robustness to Noise: The system maintained high predictive performance (AUC often >0.90) even when 90% of the features were irrelevant noise.
- Error Correction: Interestingly, the surrogate model can act as an auditor. In the Arrhythmia dataset, PROTEUS identified "True Normals"—cases where the unsupervised detector was likely wrong—at a rate of approximately 50%.
- Efficiency: Feature selection typically takes less than 2 seconds for 100-dimensional datasets, making it suitable for high-dimensional applications.
Conclusion
PROTEUS provides a robust framework for generating global, predictive explanations for unsupervised anomaly detectors. By approximating decision surfaces in low-dimensional spaces, it offers both interpretability and the ability to monitor future data. Future research will focus on extending these methods to categorical data and complex time-series dependencies.
Summary updated