2.5 Model Evaluation
Interview-ready field notes
The 20-second answer
Evaluation asks whether a model will work on unseen data and whether its errors match the real cost of being wrong.
Core ideas to retain
- Precision is ; recall is ; F1 balances them with a harmonic mean.
- Accuracy can be misleading on imbalanced data: a 99% negative-only classifier may find no positives.
- ROC-AUC measures ranking across thresholds; PR-AUC is often more informative when positives are rare.
- Cross-validation repeats train/validation splits. Use stratification for imbalanced classification and time-aware splits for time series.
Interview / OA rule
Pick metrics from error cost: prioritize recall when missing a positive is expensive; precision when false alarms are expensive. Inspect slices, not only one global score.
One good written resource
Google — Accuracy, Precision, Recall, and F1 — read this after the video when you want a clearer mental model, not more pages of notes.
Most asked
Interview questions to practise aloud
Each answer is the level of detail expected for a strong fundamentals round.
Precision vs recall?
Precision asks whether positive predictions were correct; recall asks whether actual positives were found. A threshold usually trades one for the other.
ROC-AUC vs PR-AUC?
ROC-AUC works across thresholds and is broadly useful; PR-AUC focuses on positive predictions and is usually clearer with rare positive classes.
What is data leakage?
Information unavailable at prediction time—or shared across splits—reaches training. It creates unrealistically high offline scores and production failure.