Scikit-learn

Scikit-learn is the gold standard for “Classical ML” (Regression, SVM, Random Forest). v1.6 (2025) adds Array API support (running on GPUs via PyTorch/CuPy).

When to Use

• Tabular Data: Random Forests / Gradient Boosting.
• Preprocessing: StandardScaler, LabelEncoder.
• Small Data: When Deep Learning is overkill.

Core Concepts

Estimators

Everything implements .fit(X, y) and .predict(X).

Pipelines

Chaining preprocessing and modeling: Pipeline([('scaler', StandardScaler()), ('svc', SVC())]).

Array API

Passing PyTorch tensors directly to Scikit-learn without converting to NumPy (keeping data on GPU).

Best Practices (2025)

Do:

• Use Pipelines: Prevent data leakage during cross-validation.
• Use HistGradientBoostingClassifier: It is much faster than standard extraction implementation (inspired by LightGBM).

Don’t:

• Don’t use for Images/Audio: Use PyTorch/DL for unstructured data.

References

• Scikit-learn Documentation