[python-package] Correctly recognize LGBMClassifier(num_class=2, objective="multiclass") as multiclass classification (#6524)

python-package/lightgbm/basic.py

@@ -157,6 +157,8 @@ _LGBM_SetFieldType = Union[
 
 ZERO_THRESHOLD = 1e-35
 
+_MULTICLASS_OBJECTIVES = {"multiclass", "multiclassova", "multiclass_ova", "ova", "ovr", "softmax"}
+
 
 def _is_zero(x: float) -> bool:
     return -ZERO_THRESHOLD <= x <= ZERO_THRESHOLD

python-package/lightgbm/sklearn.py

@@ -10,6 +10,7 @@ import numpy as np
 import scipy.sparse
 
 from .basic import (
+    _MULTICLASS_OBJECTIVES,
     Booster,
     Dataset,
     LightGBMError,
@@ -467,7 +468,7 @@ def _extract_evaluation_meta_data(
         # It's possible, for example, to pass 3 eval sets through `eval_set`,
         # but only 1 init_score through `eval_init_score`.
         #
-        # This if-else accounts for that possiblity.
+        # This if-else accounts for that possibility.
         if len(collection) > i:
             return collection[i]
         else:
@@ -1011,7 +1012,7 @@ class LGBMModel(_LGBMModelBase):
                 f"match the input. Model n_features_ is {self._n_features} and "
                 f"input n_features is {n_features}"
             )
-        # retrive original params that possibly can be used in both training and prediction
+        # retrieve original params that possibly can be used in both training and prediction
         # and then overwrite them (considering aliases) with params that were passed directly in prediction
         predict_params = self._process_params(stage="predict")
         for alias in _ConfigAliases.get_by_alias(
@@ -1251,7 +1252,7 @@ class LGBMClassifier(_LGBMClassifierBase, LGBMModel):
                 eval_metric_list = [eval_metric]
             else:
                 eval_metric_list = []
-            if self._n_classes > 2:
+            if self.__is_multiclass:
                 for index, metric in enumerate(eval_metric_list):
                     if metric in {"logloss", "binary_logloss"}:
                         eval_metric_list[index] = "multi_logloss"
@@ -1361,7 +1362,7 @@ class LGBMClassifier(_LGBMClassifierBase, LGBMModel):
                 "Returning raw scores instead."
             )
             return result
-        elif self._n_classes > 2 or raw_score or pred_leaf or pred_contrib:  # type: ignore [operator]
+        elif self.__is_multiclass or raw_score or pred_leaf or pred_contrib:  # type: ignore [operator]
             return result
         else:
             return np.vstack((1.0 - result, result)).transpose()
@@ -1389,6 +1390,11 @@ class LGBMClassifier(_LGBMClassifierBase, LGBMModel):
             raise LGBMNotFittedError("No classes found. Need to call fit beforehand.")
         return self._n_classes
 
+    @property
+    def __is_multiclass(self) -> bool:
+        """:obj:`bool`:  Indicator of whether the classifier is used for multiclass."""
+        return self._n_classes > 2 or (isinstance(self._objective, str) and self._objective in _MULTICLASS_OBJECTIVES)
+
 
 class LGBMRanker(LGBMModel):
     """LightGBM ranker.

tests/python_package_test/test_sklearn.py

@@ -719,6 +719,25 @@ def test_predict():
     with pytest.raises(AssertionError):
         np.testing.assert_allclose(res_engine, res_sklearn_params)
 
+    # Test multiclass binary classification
+    num_samples = 100
+    num_classes = 2
+    X_train = np.linspace(start=0, stop=10, num=num_samples * 3).reshape(num_samples, 3)
+    y_train = np.concatenate([np.zeros(int(num_samples / 2 - 10)), np.ones(int(num_samples / 2 + 10))])
+
+    gbm = lgb.train({"objective": "multiclass", "num_class": num_classes, "verbose": -1}, lgb.Dataset(X_train, y_train))
+    clf = lgb.LGBMClassifier(objective="multiclass", num_classes=num_classes).fit(X_train, y_train)
+
+    res_engine = gbm.predict(X_train)
+    res_sklearn = clf.predict_proba(X_train)
+
+    assert res_engine.shape == (num_samples, num_classes)
+    assert res_sklearn.shape == (num_samples, num_classes)
+    np.testing.assert_allclose(res_engine, res_sklearn)
+
+    res_class_sklearn = clf.predict(X_train)
+    np.testing.assert_allclose(res_class_sklearn, y_train)
+
 
 def test_predict_with_params_from_init():
     X, y = load_iris(return_X_y=True)
@@ -1035,6 +1054,20 @@ def test_metrics():
     assert len(gbm.evals_result_["training"]) == 1
     assert "binary_logloss" in gbm.evals_result_["training"]
 
+    # the evaluation metric changes to multiclass metric even num classes is 2 for multiclass objective
+    gbm = lgb.LGBMClassifier(objective="multiclass", num_classes=2, **params).fit(
+        eval_metric="binary_logloss", **params_fit
+    )
+    assert len(gbm._evals_result["training"]) == 1
+    assert "multi_logloss" in gbm.evals_result_["training"]
+
+    # the evaluation metric changes to multiclass metric even num classes is 2 for ovr objective
+    gbm = lgb.LGBMClassifier(objective="ovr", num_classes=2, **params).fit(eval_metric="binary_error", **params_fit)
+    assert gbm.objective_ == "ovr"
+    assert len(gbm.evals_result_["training"]) == 2
+    assert "multi_logloss" in gbm.evals_result_["training"]
+    assert "multi_error" in gbm.evals_result_["training"]
+
 
 def test_multiple_eval_metrics():
     X, y = load_breast_cancer(return_X_y=True)