ML Foundations

• Supervised: learn X → y from labelled examples
• Regression = continuous value; classification = category
• Unsupervised: structure in unlabelled X only
• Clustering (k-means), PCA, anomaly detection
• RL: agent learns a policy maximising cumulative reward
• No labelled answers; feedback is delayed

• House price? Regression: the target is continuous
• RL vs supervised? RL gets delayed rewards, not a label per example
• Clustering accuracy? No: there are no true labels to compare against

• Features = input columns X
• Label (target) = column y the model predicts
• One row = one example

• Typical split: 60/20/20 or 70/15/15
• Imbalanced classes: stratified split
• Time series: time-ordered split, never random

• Scaler, imputer or feature selection fitted on full data leaks test information
• Fit on train only, then transform val/test: use a Pipeline

from sklearn.pipeline import make_pipeline
pipe = make_pipeline(StandardScaler(), SVC())
pipe.fit(X_train, y_train)  # scaler sees train only
pipe.score(X_val, y_val)  # reuses train mean/std

preprocessing lives inside the pipeline

• Why validation and test? Tuning on validation makes it optimistic; untouched test gives the honest score
• Data leakage? Test information reaching training; fit preprocessing on train only, inside a Pipeline
• Time-series split? Order by time: train on the past, test on the future

• Plot error vs training size, or error vs model complexity
• Val error bottoms out, then rises as complexity grows
• The train-val gap is the overfit symptom

• 99% train, 70% test? Overfit (large gap): add data, regularise, simplify, early stop
• Early stopping? Halts training once val error starts rising, before noise is memorised
• More data fixes underfitting? No: the model lacks capacity, so add capacity or features

• Expected error = bias² + variance + irreducible noise
• Noise is in the data: no model can remove it

• Bias vs variance? Line on curved data misses it (bias); deep tree changes per sample (variance)
• Why a trade-off? Complexity cuts bias but raises variance, so total error is U-shaped
• Which ensemble? Bagging cuts variance; boosting cuts bias

• Usual choice: k=5 or 10
• cross_val_score(model, X, y, cv=5) returns k scores

• GridSearchCV uses CV for model selection and tuning
• Keep an untouched test set for the final score
• Leave-one-out: costly, high variance

• Preprocessing outside the Pipeline (leakage)
• Tuning on the test set
• Reporting only the mean, ignoring the std across folds

• Why CV, not one split? Every row validates once; the mean is steadier.
• When is plain k-fold wrong? Imbalanced classes, time data, grouped rows.
• Nested CV? Inner loop tunes, outer loop scores; avoids optimistic bias.

• Stratify for classes; order by time for time series
• Fit preprocessing on train only, then transform val/test

• Workflow? Split, fit pipeline on train, tune on val or CV, test once.
• Pick a model? Compare CV mean and std; prefer the simpler on a tie; confirm once on test.
• Cut variance? More data, regularisation, bagging or early stopping.