Conformalization Strategies¶
Calibration strategies for conformal anomaly detection with different trade-offs between computational efficiency and statistical robustness.
Available Strategies¶
Split Strategy¶
Simple train/calibration split for fast, straightforward conformal prediction.
from nonconform.strategy.split import Split
# Use 30% of data for calibration
strategy = Split(n_calib=0.3)
# Use fixed number of samples for calibration
strategy = Split(n_calib=100)
Characteristics: - Fastest computation - Simplest implementation - Least robust for small datasets - Memory efficient
Cross-Validation Strategy¶
K-fold cross-validation for robust calibration using all data.
from nonconform.strategy.cross_val import CrossValidation
# 5-fold cross-validation
strategy = CrossValidation(n_folds=5, plus=False)
# Enable plus mode for tighter prediction intervals
strategy = CrossValidation(n_folds=5, plus=True)
Characteristics: - Most robust calibration - Uses all data for both training and calibration - Higher computational cost - Recommended for small datasets
Bootstrap Strategy¶
Bootstrap resampling with configurable ensemble parameters.
from nonconform.strategy.bootstrap import Bootstrap
# Basic bootstrap with 100 models
strategy = Bootstrap(n_bootstraps=100, resampling_ratio=0.8)
# Automated configuration
strategy = Bootstrap(n_calib=200) # Auto-calculate other parameters
Characteristics: - Flexible ensemble size - Uncertainty quantification - Robust to outliers - Configurable computational cost
Jackknife Strategy¶
Leave-one-out cross-validation for maximum data utilization.
from nonconform.strategy.jackknife import Jackknife
# Standard jackknife
strategy = Jackknife(plus=False)
# Jackknife+ for tighter intervals
strategy = Jackknife(plus=True)
Characteristics: - Maximum data utilization - Computationally intensive - Best for very small datasets - Provides individual sample influence
Strategy Selection Guide¶
| Dataset Size | Computational Budget | Recommendation |
|---|---|---|
| Large (>1000) | Low | Split |
| Large (>1000) | High | CrossValidation |
| Medium (100-1000) | Any | CrossValidation |
| Small (<100) | Any | Jackknife |
Plus Mode¶
All strategies support "plus" mode for tighter prediction intervals:
# Enable plus mode for any strategy
strategy = CrossValidation(n_folds=5, plus=True)
strategy = Bootstrap(n_bootstraps=50, plus=True)
strategy = Jackknife(plus=True)
Plus mode provides: - Higher statistical efficiency in theory - Better finite-sample properties - Slightly higher computational cost
Performance Comparison¶
| Strategy | Training Time | Memory Usage | Calibration Quality |
|---|---|---|---|
| Split | Fast | Low | Good |
| CrossValidation | Medium | Medium | Excellent |
| Bootstrap | Medium-High | Medium-High | Very Good |
| Jackknife | Slow | High | Excellent |
Integration with Detectors¶
All strategies work with any conformal detector:
from nonconform.estimation.standard import ConformalDetector
from nonconform.estimation.weighted import ConformalDetector
from pyod.models.lof import LOF
# Standard conformal with cross-validation
detector = ConformalDetector(
detector=LOF(),
strategy=CrossValidation(k=5)
)
# Weighted conformal with bootstrap
detector = ConformalDetector(
detector=LOF(),
strategy=Bootstrap(n_bootstraps=100)
)