nonconform 0.9.0

Conformal Anomaly Detection

nonconform is a Python library that enhances anomaly detection by providing uncertainty quantification. It acts as a wrapper around most detectors from the popular PyOD library (see Supported Estimators). By leveraging one-class classification principles and conformal inference, nonconform enables statistically rigorous anomaly detection.

Key Features

• Uncertainty Quantification: Go beyond simple anomaly scores; get statistically valid p-values.
• Error Control: Reliably control metrics like the False Discovery Rate (FDR).
• Broad PyOD Compatibility: Works with a wide range of PyOD estimators (see Supported Estimators).
• Flexible Strategies: Implements various conformal strategies like Split-Conformal and Bootstrap-after-Jackknife+ (JaB+).

:hatching_chick: Getting Started

pip install nonconform

For additional features, you might need optional dependencies:

• pip install nonconform[data] - Includes pyarrow for loading example data (via remote download)
• pip install nonconform[dl] - Includes deep learning dependencies (TensorFlow, PyTorch)
• pip install nonconform[all] - Includes all optional dependencies

Please refer to the pyproject.toml for details.

Split-Conformal (also Inductive) Approach

Using a Gaussian Mixture Model on the Shuttle dataset:

Note: The examples below use the built-in datasets. Install with pip install nonconform[data] to run these examples.

from pyod.models.gmm import GMM
from scipy.stats import false_discovery_control

from nonconform.strategy import Split
from nonconform.estimation import StandardConformalDetector
from nonconform.utils.data import load_shuttle
from nonconform.utils.stat import false_discovery_rate, statistical_power

x_train, x_test, y_test = load_shuttle(setup=True)

ce = StandardConformalDetector(
    detector=GMM(),
    strategy=Split(calib_size=1_000)
)

ce.fit(x_train)
estimates = ce.predict(x_test)

decisions = false_discovery_control(estimates, method='bh') <= 0.2

print(f"Empirical FDR: {false_discovery_rate(y=y_test, y_hat=decisions)}")
print(f"Empirical Power: {statistical_power(y=y_test, y_hat=decisions)}")

Output:

Empirical FDR: 0.108
Empirical Power: 0.99

:hatched_chick: Advanced Usage

Bootstrap-after-Jackknife+ (JaB+)

The BootstrapConformal() strategy allows to set 2 of the 3 parameters resampling_ratio, n_boostraps and n_calib. For either combination, the remaining parameter will be filled automatically. This allows exact control of the calibration procedure when using a bootstrap strategy.

from pyod.models.iforest import IForest
from scipy.stats import false_discovery_control

from nonconform.estimation import StandardConformalDetector
from nonconform.strategy import Bootstrap
from nonconform.utils.data import load_shuttle
from nonconform.utils.stat import false_discovery_rate, statistical_power

x_train, x_test, y_test = load_shuttle(setup=True)

ce = StandardConformalDetector(
    detector=IForest(behaviour="new"),
    strategy=Bootstrap(resampling_ratio=0.99, n_bootstraps=20, plus=True)
)

ce.fit(x_train)
estimates = ce.predict(x_test)

decisions = false_discovery_control(estimates, method='bh') <= 0.1

print(f"Empirical FDR: {false_discovery_rate(y=y_test, y_hat=decisions)}")
print(f"Empirical Power: {statistical_power(y=y_test, y_hat=decisions)}")

Output:

Empirical FDR: 0.067
Empirical Power: 0.98

Weighted Conformal Anomaly Detection

The statistical validity of conformal anomaly detection depends on data exchangability (weaker than i.i.d.). This assumption can be slightly relaxed by computing weighted conformal p-values.

from pyod.models.iforest import IForest
from scipy.stats import false_discovery_control

from nonconform.utils.data import load_shuttle
from nonconform.estimation import WeightedConformalDetector
from nonconform.strategy import Split
from nonconform.utils.stat import false_discovery_rate, statistical_power

x_train, x_test, y_test = load_shuttle(setup=True)

model = IForest(behaviour="new")
strategy = Split(calib_size=1_000)

ce = WeightedConformalDetector(detector=model, strategy=strategy)
ce.fit(x_train)
estimates = ce.predict(x_test)

decisions = false_discovery_control(estimates, method='bh') <= 0.1

print(f"Empirical FDR: {false_discovery_rate(y=y_test, y_hat=decisions)}")
print(f"Empirical Power: {statistical_power(y=y_test, y_hat=decisions)}")

Output:

Empirical FDR: 0.077
Empirical Power: 0.96

Citation

If you find this repository useful for your research, please cite following papers:

Leave-One-Out-, Bootstrap- and Cross-Conformal Anomaly Detectors

@inproceedings{Hennhofer2024,
	title        = {{ Leave-One-Out-, Bootstrap- and Cross-Conformal Anomaly Detectors }},
	author       = {Hennhofer, Oliver and Preisach, Christine},
	year         = 2024,
	month        = {Dec},
	booktitle    = {2024 IEEE International Conference on Knowledge Graph (ICKG)},
	publisher    = {IEEE Computer Society},
	address      = {Los Alamitos, CA, USA},
	pages        = {110--119},
	doi          = {10.1109/ICKG63256.2024.00022},
	url          = {https://doi.ieeecomputersociety.org/10.1109/ICKG63256.2024.00022}
}

Testing for outliers with conformal p-values

@article{Bates2023,
	title        = {Testing for outliers with conformal p-values},
	author       = {Bates,  Stephen and Candès,  Emmanuel and Lei,  Lihua and Romano,  Yaniv and Sesia,  Matteo},
	year         = 2023,
	month        = feb,
	journal      = {The Annals of Statistics},
	publisher    = {Institute of Mathematical Statistics},
	volume       = 51,
	number       = 1,
	doi          = {10.1214/22-aos2244},
	issn         = {0090-5364},
	url          = {http://dx.doi.org/10.1214/22-AOS2244}
}

Model-free selective inference under covariate shift via weighted conformal p-values

@inproceedings{Jin2023,
	title        = {Model-free selective inference under covariate shift via weighted conformal p-values},
	author       = {Ying Jin and Emmanuel J. Cand{\`e}s},
	year         = 2023,
	url          = {https://api.semanticscholar.org/CorpusID:259950903}
}

Supported Estimators

The package only supports anomaly estimators that are suitable for unsupervised one-class classification. As respective detectors are therefore exclusively fitted on normal (or non-anomalous) data, parameters like threshold are internally set to the smallest possible values.

Models that are currently supported include:

• Angle-Based Outlier Detection (ABOD)
• Autoencoder (AE)
• Cook's Distance (CD)
• Copula-based Outlier Detector (COPOD)
• Deep Isolation Forest (DIF)
• Empirical-Cumulative-distribution-based Outlier Detection (ECOD)
• Gaussian Mixture Model (GMM)
• Histogram-based Outlier Detection (HBOS)
• Isolation-based Anomaly Detection using Nearest-Neighbor Ensembles (INNE)
• Isolation Forest (IForest)
• Kernel Density Estimation (KDE)
• k-Nearest Neighbor (kNN)
• Kernel Principal Component Analysis (KPCA)
• Linear Model Deviation-base Outlier Detection (LMDD)
• Local Outlier Factor (LOF)
• Local Correlation Integral (LOCI)
• Lightweight Online Detector of Anomalies (LODA)
• Locally Selective Combination of Parallel Outlier Ensembles (LSCP)
• GNN-based Anomaly Detection Method (LUNAR)
• Median Absolute Deviation (MAD)
• Minimum Covariance Determinant (MCD)
• One-Class SVM (OCSVM)
• Principal Component Analysis (PCA)
• Quasi-Monte Carlo Discrepancy Outlier Detection (QMCD)
• Rotation-based Outlier Detection (ROD)
• Subspace Outlier Detection (SOD)
• Scalable Unsupervised Outlier Detection (SUOD)

Contact

Bug reporting: https://github.com/OliverHennhoefer/nonconform/issues