Examples¶
Practical examples demonstrating different conformal anomaly detection approaches.
Getting Started¶
| Example | Difficulty | What You'll Learn |
|---|---|---|
| Classical Conformal | Beginner | Basic split conformal detection, FDR control, p-value interpretation |
| Cross-Validation Conformal | Beginner | K-fold cross-validation for better data efficiency |
Advanced Strategies¶
| Example | Difficulty | What You'll Learn |
|---|---|---|
| Jackknife Conformal | Intermediate | Jackknife+ strategy for theoretical guarantees |
| Bootstrap Conformal | Intermediate | JaB+ strategy for uncertainty quantification |
| Conditional Conformal | Intermediate | Conditionally calibrated conformal p-values with BH-style FDR selection |
Special Topics¶
| Example | Difficulty | What You'll Learn |
|---|---|---|
| Weighted Conformal | Advanced | Handling distribution shift between training and test data |
| FDR Control | Intermediate | Multiple testing correction, Benjamini-Hochberg procedure |
What Each Example Covers¶
Classical Conformal - Start here if you're new to nonconform. Learn the core workflow: wrap a detector, compute p-values, and apply FDR control. Includes visualization of results.
Cross-Validation Conformal - Use all your training data for both fitting and calibration using k-fold CV. Good for smaller datasets where you can't afford to reserve data.
Jackknife Conformal - Apply the Jackknife+ strategy, which provides finite-sample theoretical guarantees. A good balance between computational cost and accuracy.
Bootstrap Conformal - Use the Jackknife+-after-Bootstrap (JaB+) strategy for robust uncertainty quantification. Best when you need to understand detection stability.
Conditional Conformal - Apply conditional calibration maps (mc, simes, dkwm, asymptotic) to empirical conformal p-values, then run BH-style FDR selection.
Weighted Conformal - Handle covariate shift scenarios where your test data comes from a different distribution than your training data. Essential for real-world deployment.
FDR Control - Deep dive into False Discovery Rate control. Understand when to use BH vs weighted methods and how to evaluate FDR performance.
Prerequisites¶
All examples assume you have installed nonconform with the PyOD and data extras:
Most examples use the PyOD LOF detector and benchmark datasets from oddball. Each example is self-contained and can be run independently.