NormalcyScore

Dealing with Uncertainty in Contextual Anomaly Detection [TMLR 2026]

Implementation of the normalcy score (NS) framework presented in Dealing with Uncertainty in Contextual Anomaly Detection.

NS is a contextual anomaly detection method that models both aleatoric and epistemic uncertainty via heteroscedastic Gaussian process regression (HGPR), returning:

• an anomaly score (expected NS)
• an uncertainty estimate through a 95% Highest Density Interval (HDI) on the score

---

Libraries Used

• numpy, pandas, scipy
• scikit-learn
• tensorflow, gpflow
• arviz (HDI computation)
• pyod (optional baselines / utilities)
• statsmodels, patsy (optional utilities)
• rich (console logging)

---

Repository Structure

• datasets/
  Put your CSV datasets here.

• NS.py
  Implements the normalcy score: score definition and computation (and uncertainty-related outputs).

• test_NS.py
  Runnable script to test NS on a dataset from datasets/ and save results.

• ContextualAnomalyInject.py
  Utility to inject contextual anomalies and generate ground_truth.

The code structure and the anomaly injection script were implemented following the QCAD [DAMI 2023] implementation available here.

Citing the Paper

If you use this code, please cite:

@article{
bindini2026dealing,
title={Dealing with Uncertainty in Contextual Anomaly Detection},
author={Luca Bindini and Lorenzo Perini and Stefano Nistri and Jesse Davis and Paolo Frasconi},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=yLoXQDNwwa},
note={}
}

LICENSE

All material is available under Creative Commons BY-NC 4.0. You can use, redistribute, and adapt the material for non-commercial purposes, as long as you give appropriate credit by citing our paper and indicate any changes you've made.