Machine Learning in Python

scikit-learn is the starting point for classical machine learning on tabular data. It gives every model the same fit and predict interface, plus preprocessing, pipelines, cross-validation and metrics. Many libraries follow its API, so they plug into the same pipelines: imbalanced-learn adds resampling for skewed classes, Feature-engine adds feature transformers, and sktime applies the same style to time series.

On structured data, gradient boosting usually beats other model families. XGBoost, LightGBM and CatBoost all train boosted decision trees but differ in details: LightGBM is known for fast training on large datasets, CatBoost handles categorical features natively, and XGBoost is the most widely deployed. All three accept scikit-learn style estimators, so trying two of them on your data is cheap. Optuna and Hyperopt tune their hyperparameters, with Optuna's define-by-run API and trial pruning making it the common pick for new projects.

For forecasting, Prophet fits business series with trend, seasonality and holiday effects, StatsForecast runs classical models such as ARIMA and ETS quickly across many series, and TimesFM is a pretrained foundation model from Google Research that forecasts without per-series training. pgmpy covers Bayesian networks and causal inference. Start forecasting with a simple statistical baseline before moving to heavier models, because seasonal naive and ETS forecasts are often hard to beat and make it clear whether a more complex model is earning its cost.