Monotone Constraints#

The monotone_constraints module provides functions for inferring monotone constraints for XGBoost models.

Functions#

infer_monotone_constraints_from_correlations#

iguanas.monotone_constraints.infer_monotone_constraints_from_correlations(X: polars.DataFrame, y: polars.Series) → polars.DataFrame[source]#

Compute monotone constraint signs for XGBoost based on feature-target correlations.

Parameters:

  • X (pl.DataFrame) – DataFrame containing features.

  • y (pl.Series) – Target series for computing correlations.

Returns:

DataFrame with columns:

  • feature: Feature name

  • pearson_corr: Pearson correlation with target

  • constraint: Constraint value (1 for positive constraint, -1 for negative constraint, 0 for no constraint)

Return type:

pl.DataFrame

Compute monotone constraint signs for XGBoost based on feature-target correlations.

Calculates Pearson correlations between each feature and the target, then assigns constraint values:

  • 1 for positive correlation

  • -1 for negative correlation

  • 0 for no correlation

Parameters:

  • X (pl.DataFrame): DataFrame containing features

  • y (pl.Series): Target series for computing correlations

Returns:

  • pl.DataFrame: DataFrame with columns: feature, pearson_corr, constraint

infer_monotone_constraints_from_stumps#

iguanas.monotone_constraints.infer_monotone_constraints_from_stumps(stump: xgboost.sklearn.XGBClassifier, X: polars.DataFrame, y: polars.Series) → polars.DataFrame[source]#

Determine monotone constraints by training decision stumps for each feature.

Trains a single-split tree (max_depth=1) for each feature and examines how predictions change from min to max value to determine monotonic relationship.

Parameters:

  • stump (XGBClassifier) – XGBoost classifier configured as a stump (max_depth=1).

  • X (pl.DataFrame) – Features DataFrame.

  • y (pl.Series) – Target series for training.

Returns:

DataFrame with columns: - feature: Feature name - constraint: Constraint value (1 for positive constraint, -1 for negative constraint, 0 for no constraint) - pred_at_min: Predicted probability at minimum feature value - pred_at_max: Predicted probability at maximum feature value - delta: Change in probability (pred_at_max - pred_at_min)

Return type:

pl.DataFrame

Determine monotone constraints by training decision stumps for each feature.

Trains a single-split tree (max_depth=1) for each feature and examines how predictions change from min to max value to determine monotonic relationship.

Parameters:

  • stump (XGBClassifier): XGBoost classifier configured as a stump (max_depth=1)

  • X (pl.DataFrame): Features DataFrame

  • y (pl.Series): Target series for training

Returns:

  • pl.DataFrame: DataFrame with columns: feature, constraint, pred_at_min, pred_at_max, delta