In the LightGBM documentation it is stated that one can set predict_contrib=True to predict the SHAP-values.
How do we extract the SHAP-values (apart from using the shap package)?
I have tried
model = LGBM(objective="binary",is_unbalance=True,predict_contrib=True)
model.fit(X_train,y_train)
pred_shap = opt_model.predict(X_train) #Does not get SHAP-values
which does not seem to work
Shap values the LGBM way with pred_contrib=True:
from lightgbm.sklearn import LGBMClassifier
from sklearn.datasets import load_iris
X,y = load_iris(return_X_y=True)
lgbm = LGBMClassifier()
lgbm.fit(X,y)
lgbm_shap = lgbm.predict(X, pred_contrib=True)
# Shape of returned LGBM shap values: 4 features x 3 classes + 3 expected values over the training dataset
print(lgbm_shap.shape)
# 0th row of LGBM shap values for 0th feature
print(lgbm_shap[0,:4])
Output:
(150, 15)
[-0.0176954 0.50644615 5.56584344 3.43032313]
Shap values from shap:
import shap
explainer = shap.TreeExplainer(lgbm)
shap_values = explainer.shap_values(X)
# num of predicted classes
print(len(shap_values))
# shap values for 0th class for 0th row
print(shap_values[0][0])
Output:
3
array([-0.0176954 , 0.50644615, 5.56584344, 3.43032313])
Looks the same to me.
The confusion arises because of the duplication (inconsistent naming) of the controlling parameter in two different lightgbm APIs.
Each of the two major APIs uses its own spelling:
And the documentation favors the C version (the python API spelling is not even considered an alias...)