Reproducible error: SHAP ExplainerError: Additivity check failed in TreeExplainer

This is related to issue #866

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler

# Create a sample DataFrame X
np.random.seed(0)  # for reproducibility
data = {
    'Category': np.random.randint(2, size=100),
    'Value': np.random.randint(100, size=100)
}
X = pd.DataFrame(data)

# Creating binary arrays T and y
T = np.random.randint(2, size=100)
y = np.random.randint(2, size=100)

# Splitting the data into training and testing sets
X_train, X_test, T_train, T_test, y_train, y_test = train_test_split(X, T, y, test_size=0.2, random_state=42)

to_scale = ["Value"]
scaler = MinMaxScaler()
# Fit scaler on training data and transform both training and testing data
X_train[to_scale] = scaler.fit_transform(X_train[to_scale])
X_test[to_scale] = scaler.transform(X_test[to_scale])

est = CausalForestDML(model_y=RandomForestClassifier(), model_t=DummyClassifier(strategy='uniform'), random_state=123, discrete_outcome=True,discrete_treatment=True)
est.fit(y_train, T_train, X=X_train)
shap_values = est.shap_values(X_train[:20])
shap.plots.beeswarm(shap_values['Y0']['T0_1'])

If you try to comment out or remove the lines:

to_scale = ["Value"]
scaler = MinMaxScaler()
# Fit scaler on training data and transform both training and testing data
X_train[to_scale] = scaler.fit_transform(X_train[to_scale])
X_test[to_scale] = scaler.transform(X_test[to_scale])

the code runs correctly.

It seems that the scaler induces an issue, it is the same with StandardScaler(). My guess is that float rounding errors are responsible for this behavior. I believe that there is a PR to add check_additivity = False, which seems to be the only way to resolve this, unless I am missing something.

Start by running the supplied CausalForestDML and TreeExplainer reproducer with and without MinMaxScaler or StandardScaler, then compare the resulting additivity-check failure. Review the related issue #866 and the referenced check_additivity change; done means the scaled example behaves correctly or the limitation and required workaround are clearly established with regression coverage.