scikit-learn

For tabular data and classical statistics, most Python practitioners still reach for scikit-learn first. It bundles linear and logistic models, support vector machines, nearest neighbors, decision trees, random forests, histogram gradient boosting, clustering, PCA and other decompositions, plus cross-validation, hyperparameter search, and evaluation metrics, all exposed through the same fit, predict, and transform methods so individual steps chain into Pipeline objects. Release 1.9.1, from September 2026, needs Python 3.11 or newer plus NumPy, SciPy, joblib, threadpoolctl, and Narwhals, and installs from PyPI or conda-forge. Computation is CPU based and parallelized through joblib, though an experimental array API mode lets selected estimators run on PyTorch, CuPy, or dpnp arrays, GPUs included. Fitted models can be saved with pickle or with the skops format, which is safer to load. The BSD 3-Clause license puts almost no limits on commercial use. Run by volunteers with institutional backing, it suits analysts, data scientists, and engineers building baselines or production models on structured data.