the random forest regressor with sklearnex patch produce much larger MSE than the original sklearn random forest regressor.

[zhuang-hao-ming]

Describe the bug

the random forest regressor with sklearnex patch produce much larger MSE than the original sklearn random forest regressor.

The example in the document, compares implementations with/without patch by calling patch_sklearn()/unpatch_sklearn(), which produce similar MSE. However, use sklearn directly produce much less MSE.

To Reproduce

remove patch_sklearn()/unpatch_sklearn() in the random forest example in the document can produce the error

Expected behavior Describe what your are expecting from steps above

Output/Screenshots If applicable, add output/screenshots to help explain your problem.

Environment:

• OS: windows
• Version: 2021

[napetrov]

Can you please share more details on dataset where you observe this difference?

[zhuang-hao-ming]

thank you very much for your response. I have shared the dataset that can observed this difference in the google driver

https://drive.google.com/file/d/1FGDFcUM1-XnIUCt3jXQ9hGANQFF2vIeU/view?usp=sharing

the code for producing this difference are shown below. the one with sklearnex patch get mse (67569.56223138234), while the one without get mse (20874.257079172243).

x_train, y_train, x_test, y_test = np.load('./data.npz')['x_train'], np.load('./data.npz')['y_train'], np.load('./data.npz')['x_test'], np.load('./data.npz')['y_test']

# from sklearnex import patch_sklearn  # patch or not
# patch_sklearn()

from sklearn.ensemble import RandomForestRegressor
params = {
    'n_estimators': 150,
    'random_state': 44,
    'n_jobs': -1,
    'random_state': 1
}
start = timer()
rf = RandomForestRegressor(**params).fit(x_train, y_train)
train_patched = timer() - start
print(f"Intel® extension for Scikit-learn time: {train_patched:.2f} s")
y_pred = rf.predict(x_test)
mse_opt = metrics.mean_squared_error(y_test, y_pred)
print(f'Intel® extension for Scikit-learn Mean Squared Error: {mse_opt}')

[ahuber21]

Hi @zhuang-hao-ming thank you for reporting this issue. I will start working on this today and firstly will try to reproduce your observations. I have just opened a PR in the oneDAL repo that will change how we sample our features on node splits https://github.com/oneapi-src/oneDAL/pull/2292 which could have an impact here. I will update you once I know more.