Neural regression checkpoint · Data Playground

Outputs you create

search: Fitted GridSearchCV for the declared parameter comparison.
cv_results: Candidate validation evidence; comparison workflows use a dataframe indexed by model ID.
reference_results: cross_validate result for the dummy reference; test_score contains five scores.
loss_curve: Training loss history of the selected neural regressor.
residuals: Training-only actual, predicted and actual-minus-predicted residual evidence.
final_rmse: Root mean squared error in original target units.

Wine600 · first 8 rows
`fixed acidity`	`volatile acidity`	`citric acid`	`residual sugar`	`chlorides`	`free sulfur dioxide`	`total sulfur dioxide`	`density`	`pH`	`sulphates`	`alcohol`	`quality`	`wine_type`
5.3	0.32	0.12	6.6	0.043	22	141	0.9937	3.36	0.6	10.4	6	white
6.4	0.3	0.27	4.4	0.055	17	135	0.9925	3.23	0.44	12.2	6	white
6.5	0.22	0.32	2.2	0.028	36	92	0.99076	3.27	0.59	11.9	7	white
8	0.24	0.48	6.8	0.047	13	134	0.99616	3.23	0.7	10	5	white
8.9	0.29	0.35	1.9	0.067	25	57	0.997	3.18	1.36	10.3	6	red
6.6	0.21	0.39	2.3	0.041	31	102	0.99221	3.22	0.58	10.9	7	white
8.8	0.24	0.54	2.5	0.083	25	57	0.9983	3.39	0.54	9.2	5	red
5.4	0.15	0.32	2.5	0.037	10	51	0.98878	3.04	0.58	12.6	6	white

Your task

Complete the Wine600 neural regression workflow with feature/target scaling, width selection, convergence evidence and original-unit errors.

searchcv_resultsreference_resultsloss_curveresidualsfinal_rmse

Hints

Think: Reconstruct both feature and target preparation inside every neural fit.

Tools: Pipeline, TransformedTargetRegressor, nested width search, loss curve and original-unit RMSE.

Approach: Compare against the mean reference, select width on training folds, inspect convergence and residuals, then predict through the fitted wrapper.

Explained solution

from sklearn.base import clone
from sklearn.model_selection import train_test_split, KFold, StratifiedKFold, cross_validate, cross_val_predict, GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer, TransformedTargetRegressor
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.dummy import DummyRegressor, DummyClassifier
from sklearn.metrics import root_mean_squared_error, f1_score, accuracy_score, confusion_matrix

from sklearn.neural_network import MLPRegressor
X = df.drop(columns=['quality', 'wine_type'])
y = df['quality']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
model = Pipeline([('prepare', StandardScaler()), ('model', TransformedTargetRegressor(regressor=MLPRegressor(hidden_layer_sizes=(24,), max_iter=800, early_stopping=True, tol=1e-3, random_state=42), transformer=StandardScaler()))])
folds = KFold(n_splits=5, shuffle=True, random_state=42)
cv_results = cross_validate(model, X_train, y_train, cv=folds, scoring='neg_root_mean_squared_error')
reference_results = cross_validate(DummyRegressor(strategy='mean'), X_train, y_train, cv=folds, scoring='neg_root_mean_squared_error')
search = GridSearchCV(model, {'model__regressor__hidden_layer_sizes':[(16,), (24,)]}, cv=folds, scoring='neg_root_mean_squared_error')
search.fit(X_train, y_train)
oof_predictions = cross_val_predict(search.best_estimator_, X_train, y_train, cv=folds)
residuals = y_train - oof_predictions
loss_curve = search.best_estimator_.named_steps['model'].regressor_.loss_curve_
final_model = clone(search.best_estimator_).fit(X_train, y_train)
final_predictions = final_model.predict(X_test)
final_rmse = root_mean_squared_error(y_test, final_predictions)
print('Final RMSE in quality-score units:', final_rmse)

Fold-local target scaling supports optimisation without leakage, and the wrapper reverses it so final errors retain quality-score units.