The prediction of XGBoostClassifier doesn't match the output of ShapIQ

[Linh-nk]

The log odd of f(x) in the waterfall plot does not match the prediction I got from XGBoostClassifier model predict_proba method.

Also, should the baseline value be close to the log odd of the average of y? The log odd of E[f(X)] I got from the waterfall plot does not match the average of ground truth. Please correct me if I'm wrong.

[mmschlk]

I don't see an image/waterfall plot. So I can only speculate. But yes, you are right, the SV do not match with what comes out of XGBoost when you run predict_proba. Similar to SHAP, we explain the margin of the prediction and not the predicted probabilities (see this test case test for a reference):

# explain with shapiq
explainer_shapiq = TreeExplainer(
   model=xgb_clf_model, max_order=1, index="SV", class_index=class_label
)
sv = explainer_shapiq.explain(x=x_explain_shapiq)
sum_of_sv = sv

# get the margin prediction of the model
prediction = xgb_clf_model.predict(x_explain.reshape(1, -1), output_margin=True)
prediction = prediction[0, class_label]

print(prediction == sum_of_sv)
True

There is a paper discussing that the Margins are not a very nice object to explain models with. But up to now neither shap nor shapiq supports another means of explanation here.

Does this answer your question?

[Linh-nk]

Thank you for the prompt response! However, I'm still not getting the same output from shapiq and XGBClassifier. Here is the code that I used

import shapiq

explainer_shapiq = shapiq.TreeExplainer(
   model=model, max_order=1, index="SV"
)
sv = explainer_shapiq.explain(x=X.iloc[0].to_numpy())
sum_of_sv = sv.get_n_order_values(1).sum()
base_value = sv.baseline_value

prediction = model.predict(X.iloc[0].to_numpy().reshape(1, -1), output_margin=True)

print(f'prediction: {prediction}')
print(f'shapiq prediction: {base_value + sum_of_sv}')

where model is an XGBClassifier model. And here is the output I got

prediction: [0.91437876]
shapiq prediction: 2.201628711526583

[Linh-nk]

When I use the shap package, I could get the accurate prediction

import shap

explainer_shap = shap.TreeExplainer(
   model=model, #max_order=1, index="SV"
)
sv = explainer_shap(X=X.iloc[0].to_numpy().reshape(1, -1))

prediction = model.predict(X.iloc[0].to_numpy().reshape(1, -1), output_margin=True)

print(f'prediction: {prediction}')
print(f'shapiq prediction: {sv.values.sum() + sv.base_values}')

prediction: [0.91437876]
shapiq prediction: [0.9143772]

At the first glance, the SHAP values output by shap and shapiq seem to match, but their baseline_value do not

[mmschlk]

Ah okay, just to make sure: First, are you certain you are explaining the correct class in the shapiq case?
Omitting the class_index in TreeExplainer might default to some other class than what you are getting predictions for. Can you check your case for all class indices?

explainer_shapiq = shapiq.TreeExplainer(
   model=model, max_order=1, index="SV", class_index=class_index
)

Second, what happens if you do not run:

sum_of_sv = sv.get_n_order_values(1).sum()
base_value = sv.baseline_value

but

sum_of_sv = sum(sv.values)

since with shapiq v1.2.3, we made sure that min_order is set to 0 for TreeExplainer which should include the baseline value inside the sv.values array.

Third is it possible to create a minimal reproducible case where shap returns the prediction and shapiq does not? Because here it is the same.

Best
Max

[Linh-nk]

Hi Max,

I'm running the same example you have but for binary case

from sklearn.datasets import make_classification, make_regression


import numpy as np
import xgboost
import copy

def background_clf_dataset() -> tuple[np.ndarray, np.ndarray]:
    """Return a simple background dataset."""
    X, y = make_classification(
        n_samples=100,
        n_features=10,
        random_state=42,
        n_classes=2,  #binary here
        n_informative=5,
        n_repeated=0,
        n_redundant=0,
    )
    return copy.deepcopy(X), copy.deepcopy(y)

def xgb_clf_model(background_clf_dataset):
    """Return a simple xgboost classification model."""

    X, y = background_clf_dataset
    model = xgboost.XGBClassifier(random_state=42, n_estimators=3)
    model.fit(X, y)
    return model

background_clf_dataset = background_clf_dataset()
xgb_clf_model = xgb_clf_model(background_clf_dataset)
background_clf_data, y = background_clf_dataset

explanation_instance = 1
class_label = 1

# the following code is used to get the shap values from the SHAP implementation
import shap
model_copy = copy.deepcopy(xgb_clf_model)
explainer_shap = shap.TreeExplainer(model=model_copy)
baseline_shap = float(explainer_shap.expected_value)

x_explain_shap = copy.deepcopy(background_clf_data[explanation_instance].reshape(1, -1))
sv_shap_all_classes = explainer_shap.shap_values(x_explain_shap)
sv_shap = sv_shap_all_classes#[0][:, class_label]

# compute with shapiq
import shapiq
explainer_shapiq = shapiq.TreeExplainer(
    model=xgb_clf_model, max_order=1, index="SV", class_index=class_label
)
x_explain_shapiq = copy.deepcopy(background_clf_data[explanation_instance])
sv_shapiq = explainer_shapiq.explain(x=x_explain_shapiq)
sv_shapiq_values = sv_shapiq.get_n_order_values(1)
baseline_shapiq = sv_shapiq.baseline_value

prediction = xgb_clf_model.predict(x_explain_shapiq.reshape(1, -1), output_margin=True)

And here is the output that I got

baseline_shap: 0.0
SHAP values: [[ 0.21501832  0.01152     0.10970242  0.12228977  0.09716809  0.02027433
   0.8166808   0.         -0.01836885 -0.22206578]]
baseline_shapiq: 0.4894029824013125
SHAPIQ values: [ 0.21501832  0.01152     0.10970243  0.12228977  0.09716809  0.02027433
  0.81668073  0.         -0.01836885 -0.22206577]
baseline_shap + SHAP values: 1.1522190570831299
baseline_shapiq + SHAPIQ values: 1.6416220384585056
prediction: [1.1416221]

[mmschlk]

Ah okay, I will take a look at this more closely! However, the Shapley values are the same for shap and shapiq, which is at least the most important thing. It might be that the baseline_value is not properly extracted from the xgboost model. I just know that this was actually not that easy. Thank you for pointing this out!

Could you let me know what's your xgboost and shapiq version?

[Linh-nk]

It's xgboost 2.1.4 and shapiq 1.2.3. Thank you!