Shared language for practical tabular modeling.
Example: one unit represented by a row. Feature: input available at prediction time. Target: outcome to predict. Label: observed target value. Prediction time: moment when a live prediction must be made. Baseline: simple rule a useful model must beat.
Supervised learning: learning from examples with labels. Classification: predicting a category. Regression: predicting a number. Estimator: scikit-learn object that learns through fit. Parameter: value learned from data. Hyperparameter: setting chosen before fitting.
Generalization: useful performance on new examples. Leakage: training with information unavailable in genuine prediction. Cross-validation: repeated fitting with rotating held-out folds. Precision: correct positives divided by predicted positives. Recall: correct positives divided by actual positives. Threshold: score cutoff that turns probabilities into decisions.
Underfitting: model fails on training and held-out data. Overfitting: model learns training specifics that do not generalize. Drift: live data or outcome relationships change from evaluated conditions. Pipeline: ordered transforms and estimator fitted as one object.