Description

When using LGBMClassifier to train a model that is stopped by early stopping and the predict_proba method is called with a value for 'num_iteration' greater than the value stored in the model's 'best_iteration_' attribute, the method appears to still use the number of iterations specified by 'best_iteration_', rather than the number passed in the parameter.

Reproducible example

from sklearn.datasets import load_breast_cancer
from lightgbm import LGBMClassifier, early_stopping
from sklearn.metrics import log_loss
X,y = load_breast_cancer(return_X_y = True)
lgb = LGBMClassifier(n_estimators=1000)
lgb.fit(X[:400], y[:400], eval_set = (X[400:],y[400:]), callbacks=[early_stopping(10)])
for i in range(lgb.best_iteration_-5, lgb.best_iteration_+5):
print(log_loss(y[400:], lgb.predict_proba(X[400:], num_iteration=i)))
print(lgb.evals_result_['valid_0']['binary_logloss'][lgb.best_iteration_-6:lgb.best_iteration_+4])

Environment info

LightGBM version 4.5