roc_auc_score differs between RandomForestClassifier GridSearchCV and explicitly coded RandomForestCLassifier

Question

Why doesn't a trained RandomForestClassifier with specific parameters match the performance of varying those parameters with a GridSearchCV?

def random_forest(X_train, y_train):
    from sklearn.ensemble import RandomForestClassifier
    from sklearn.model_selection import GridSearchCV
    from sklearn.metrics import roc_auc_score, make_scorer
    from sklearn.model_selection import train_test_split
    
    X_train, X_validate, y_train, y_validate = train_test_split(X_train, y_train, random_state=0)
    
    # various combinations of max depth and max features
    max_depth_vals = [1,2,3]
    max_features_vals = [2,3,4]
    grid_values = {'max_depth': max_depth_vals, 'max_features': max_features_vals}
    
    # build GridSearch
    clf = RandomForestClassifier(n_estimators=10)
    grid = GridSearchCV(clf, param_grid=grid_values, cv=3, scoring='roc_auc')
    grid.fit(X_train, y_train)
    y_hat_proba = grid.predict_proba(X_validate)
    print('Train Grid best parameter (max. AUC): ', grid.best_params_)
    print('Train Grid best score (AUC): ', grid.best_score_)
    print('Validation set AUC: ', roc_auc_score(y_validate, y_hat_proba[:,1]))

    
    # build RandomForest with hard coded values. AUC should be ballpark to grid search
    clf = RandomForestClassifier(max_depth=3, max_features=4, n_estimators=10)
    clf.fit(X_train, y_train)
    y_hat = clf.predict(X_validate)
    y_hat_prob = clf.predict_proba(X_validate)[:, 1]
    
    auc = roc_auc_score(y_hat, y_hat_prob)
    
    print("\nMax Depth: 3 Max Features: 4\n---------------------------------------------")
    print("auc: {}".format(auc))
    return

Results - the grid search identifies the best parameters of max_depth=3 and max_features=4 and calculates a roc_auc_score of 0.85; when I put that through the code with the reserved validation set i get an roc_auc_score of 0.84. However when I code the classifier directly with those parameters it calculates an roc_auc_score of 1.0. My understanding is that it should be in the same ballpark ~0.85 but this feels way off.

Validation set AUC:  0.8490471073563559
Grid best parameter (max. AUC):  {'max_depth': 3, 'max_features': 4}
Grid best score (AUC):  0.8599727094965482

Max Depth: 3 Max Features: 4
---------------------------------------------
auc: 1.0

I could be misunderstanding concepts, not apply techniques correctly, or even have coding issues. Thanks.

Answer 1

There are 2 issues:

Variability

To get reproducible results, specify the seed or random state wherever possible, e.g.

RandomForestClassifier(n_estimators=10, random_state=1234)

cv = StratifiedKFold(n_splits=3, random_state=1234)
GridSearchCV(clf, param_grid=grid_values, cv=cv, scoring='roc_auc')

Parameters for ROC-AUC calculation

You use the estimated labels instead of the true labels:

auc = roc_auc_score(y_hat, y_hat_prob)

Use the true labels instead:

auc = roc_auc_score(y_validate, y_hat_prob)