[Feature] add feature importance

mqboost/regressor.py

@@ -100,6 +100,7 @@ def fit(
                 custom_metric=self._MQObj.feval,
                 evals=[(_eval_set, "eval")],
             )
+        self._colnames = dataset.columns.to_list()
         self._fitted = True
 
     def predict(
@@ -121,13 +122,25 @@ def predict(
     def __predict_available(self) -> None:
         """Check if the model has been fitted before making predictions."""
         if not getattr(self, "_fitted", False):
-            raise FittingException("Fit must be executed before predict")
+            raise FittingException("Fit must be executed first.")
 
     @property
     def MQObj(self) -> MQObjective:
         """Get the MQObjective instance."""
         return self._MQObj
 
+    @property
+    def feature_importance(self) -> dict[str, float]:
+        self.__predict_available()
+        importances = {str(k): 0 for k in self._colnames}
+        if self.__is_lgb:
+            _importance = self.model.feature_importance(importance_type="gain").tolist()
+            importances.update({str(k): v for k, v in zip(self._colnames, _importance)})
+            return importances
+        else:
+            importances.update(self.model.get_score(importance_type="gain"))
+            return importances
+
     @property
     def __is_lgb(self) -> bool:
         """Check if the model is LightGBM."""

tests/test_regressor.py

@@ -177,3 +177,39 @@ def test_monotone_constraints_called_xgb(dummy_dataset_xgb):
             for k in range(len(predictions) - 1)
         ]
     )
+
+
+def test_feature_importance_before_fit_raises():
+    params = {"learning_rate": 0.1, "max_depth": 6}
+    with pytest.raises(FittingException, match="Fit must be executed first."):
+        _ = MQRegressor(params=params).feature_importance
+
+
+def test_feature_importance_after_fit(dummy_dataset_lgb):
+    params = {"learning_rate": 0.1, "max_depth": 6}
+    gbm_model = MQRegressor(params=params)
+    gbm_model.fit(dataset=dummy_dataset_lgb)
+    feature_importances = gbm_model.feature_importance
+
+    assert isinstance(
+        feature_importances, dict
+    ), "Feature importances should be a dictionary"
+    assert len(feature_importances) == len(
+        dummy_dataset_lgb.columns
+    ), "Feature importance length mismatch"
+    for feature in dummy_dataset_lgb.columns:
+        assert (
+            str(feature) in feature_importances
+        ), f"Feature {feature} not found in importance"
+
+
+def test_feature_importance_positive(dummy_dataset_lgb):
+    """Test that at least some feature importances are non-zero after training"""
+    params = {"learning_rate": 0.1, "max_depth": 6}
+    gbm_model = MQRegressor(params=params)
+    gbm_model.fit(dataset=dummy_dataset_lgb)
+    feature_importances = gbm_model.feature_importance
+
+    assert all(
+        [importance >= 0 for importance in feature_importances.values()]
+    ), "All importance should be positive."