Skip to content

Foundation PLR Documentation

Adding New Methods

Adding New Methods¶

This tutorial explains how to extend Foundation PLR with new preprocessing methods.

Adding a New Outlier Detection Method¶

Step 1: Create the Method¶

# src/anomaly_detection/my_outlier_method.py

import numpy as np
from typing import Tuple

def detect_outliers_my_method(
    signal: np.ndarray,
    threshold: float = 0.5
) -> Tuple[np.ndarray, np.ndarray]:
    """
    Detect outliers using my custom method.

    Parameters
    ----------
    signal : np.ndarray
        Input PLR signal of shape (n_timepoints,).
    threshold : float
        Detection threshold.

    Returns
    -------
    Tuple[np.ndarray, np.ndarray]
        - outlier_mask: Boolean mask where True = outlier
        - confidence: Confidence scores per timepoint
    """
    # Your implementation here
    outlier_mask = np.zeros(len(signal), dtype=bool)
    confidence = np.zeros(len(signal))

    # Example: simple threshold-based detection
    median = np.median(signal)
    mad = np.median(np.abs(signal - median))
    z_scores = np.abs(signal - median) / (mad + 1e-8)
    outlier_mask = z_scores > threshold

    return outlier_mask, z_scores

Step 2: Register the Method¶

Add to src/anomaly_detection/anomaly_detection.py:

from src.anomaly_detection.my_outlier_method import detect_outliers_my_method

OUTLIER_METHODS = {
    # ... existing methods ...
    "my_method": detect_outliers_my_method,
}

Step 3: Add Configuration¶

# configs/outlier_methods/my_method.yaml
outlier_method: my_method
outlier_params:
  threshold: 0.5

Step 4: Test¶

python -m src.classification.flow_classification \
    outlier_method=my_method

Adding a New Imputation Method¶

Step 1: Create the Method¶

# src/imputation/my_imputation_method.py

import numpy as np

def impute_my_method(
    signal: np.ndarray,
    mask: np.ndarray
) -> np.ndarray:
    """
    Impute missing values using my custom method.

    Parameters
    ----------
    signal : np.ndarray
        Input signal with missing values.
    mask : np.ndarray
        Boolean mask where True = missing/outlier.

    Returns
    -------
    np.ndarray
        Imputed signal.
    """
    imputed = signal.copy()
    # Your implementation here
    return imputed

Step 2: Register and Configure¶

Follow the same pattern as outlier detection.

Best Practices¶

Follow NumPy docstring style for automatic API docs
Add type hints for better IDE support
Write tests in tests/ directory
Update documentation in docs/user-guide/
Register in YAML - Add to configs/mlflow_registry/parameters/classification.yaml

Registry is Single Source of Truth

After adding a new method, register it in configs/mlflow_registry/parameters/classification.yaml to ensure it's counted correctly. Currently: 11 outlier methods, 8 imputation methods.

See Also¶