Traditional anomaly detection treats the world deterministically. A model learns a representation of "normal" data and flags any significant deviation—measured via reconstruction error, distance thresholds, or density estimation—as an anomaly.
However, this deterministic framing suffers from a fundamental flaw: it conflates model confusion with true anomalies. If a data point is highly noisy, or if it falls into a region where the model simply lacks training data, a traditional detector will confidently throw a false alarm. To build reliable systems, a growing body of research shifts the objective from calculating a static deviation to explicitly modeling predictive uncertainty.
The Intuition: Why Deterministic Scoring Fails
Consider a model trained to predict or reconstruct time-series data. When faced with a high reconstruction error, a traditional system issues a binary alert. But a high error can happen for three distinct reasons:
- The data point is a true anomaly (e.g., a broken sensor or a cyberattack).
- The data is inherently noisy at that specific moment (aleatoric uncertainty).
- The model has never seen data like this before, so it is guessing blindly (epistemic uncertainty).
Without a formal representation of uncertainty, human operators cannot effectively evaluate the reliability of automated alarms, leading to alarm fatigue in high-stakes scenarios (Osborne et al.; Ma et al.). As Liu et al. note, the lack of robust uncertainty quantification is the primary driver of high false-positive rates that disrupt real-world deployments.
By treating the anomaly score as a probabilistic metric rather than a static value, modern frameworks can modulate their alarms based on how certain the model is about its own environment.
Deconstructing Uncertainty: Aleatoric vs. Epistemic
To build an uncertainty-aware anomaly score, models must separate the types of uncertainty they encounter.
Capturing Aleatoric Uncertainty (Inherent Noise)
To prevent inherent data variability from triggering false positives, models must learn the expected variance of the normal data distribution. Seo et al. achieve this using a Quantile Autoencoder (QAE). Instead of predicting a single deterministic reconstruction, the model predicts the upper, lower, and median quantiles of the data distribution.
$$\text{Aleatoric Uncertainty} \propto \text{Range}(\text{Quantile}_{\text{upper}} - \text{Quantile}_{\text{lower}})$$
If a data point deviates from the median but falls within a wide quantile range, the system recognizes that the region is naturally volatile, and the anomaly score is automatically tempered.
Capturing Epistemic Uncertainty (Model Ignorance)
When a model encounters a completely novel environment, its predictions become highly uncertain due to data sparsity. Perini et al. address this by treating the anomaly metric itself as a random variable derived from posterior distributions. By calculating a highest-density interval for each assessment, the system can flag when it is "guessing." A wide interval indicates high epistemic uncertainty, allowing practitioners to distinguish between a confident anomaly detection and an unreliable prediction caused by a lack of training data.
By integrating both dimensions, models like the one proposed by Wen et al. derive an Uncertainty-Incorporated Anomaly Score (UI-AS). This score successfully differentiates true anomalies from behaviors that are merely uncertain but still within normal bounds.
[Input Data] ───► [Probabilistic Model] ───► Total Deviation (Error)
│
▼
Is the model confident?
/ \
(Yes) (No)
/ \
▼ ▼
True Detectable Anomaly Model Confusion / Noise
(High Score, High Alarm) (Score Downgraded / Filtered)
The Modulation Mechanism: Filtering False Positives
How does uncertainty actually change the final anomaly score? The core mechanism is confidence-modulated filtering. Instead of using raw error, the anomaly score is weighted inversely by the model's uncertainty.
Flaborea et al. demonstrate this principle by defining detectable anomalies—instances where the model is highly confident in its structural understanding of the pattern but still fails to reconstruct the signal correctly. When the model identifies its own prediction as highly uncertain, the system automatically downgrades the anomaly score. This prevents false positives caused by complex or noisy signals.
Similarly, in graph-based architectures, Wei et al. utilize Graph Evidential Learning (GEL) to output higher-order evidential distributions. By integrating reconstruction error with these explicit uncertainty metrics, their framework isolates noisy normal data (which exhibits high error but low uncertainty) from genuine structural anomalies.
Calibrating Scores for Human Trust
For an uncertainty-aware system to be useful, its outputs must be statistically calibrated. Standard anomaly detection often yields uncalibrated binary labels or arbitrary scores, making it difficult for users to assess the statistical significance of an outlier.
In a loss framework for calibrated anomaly detection, the authors establish a rigorous statistical foundation for this problem. By using "proper loss functions" that focus specifically on the low-density tails of a distribution, they connect anomaly scoring directly to quantile estimation, ensuring that the resulting scores carry a mathematically sound confidence interpretation.
Once calibrated, these uncertainties can be translated into intuitive metrics. For instance, He et al. convert predictive uncertainty into explicit outlier probabilities. This allows domain experts to interpret the exact likelihood of an anomaly, building trust in automated systems. As highlighted by Laaridh et al., providing this contextual confidence score is what ultimately allows human operators to separate true system irregularities from underlying measurement or alignment errors.
Algorithmic Implementations Across Domains
The philosophy of uncertainty-aware scoring has been adapted across a wide variety of underlying machine learning paradigms:
- Gaussian Processes: Osborne et al. introduce a "fault bucket" approach for Gaussian Processes (GPs) to model unknown sensor faults as a high-variance noise process. To scale this probabilistic framework to millions of samples, Domingues et al. combine deep learning with the principled uncertainty quantification of GPs via Deep Gaussian Process Autoencoders (DGP-AE).
- Feature Memory Banks: In computer vision, Lee et al. compute anomaly scores by matching target features against a memory bank. Their scoring function incorporates the mathematical 'certainty' of a feature's nearest-neighbor match, significantly enhancing the precision of anomaly localization in industrial images.
- Synthetic Data Generation: Even when generating synthetic "auxiliary" anomalies to train detectors, uncertainty acts as a gatekeeper. Perini et al. introduce the Expected Anomaly Posterior (EAP), a Bayesian framework that quantifies the utility of synthetic training examples by balancing how anomalous they appear against how realistic they are, explicitly leveraging both aleatoric and epistemic uncertainty to filter out poor training data.
Key Foundational Work
Online novelty detection on temporal sequences (2003) — contributed 1 concept elements to this direction.
Prediction and Fault Detection of Environmental Signals with Uncharacterised Faults (2012) — contributed 1 concept elements to this direction.