Core Concepts
- Bias vs variance? — underfit vs overfit; diagnose via train/val gap.
- Overfitting fixes? — more data, regularisation, CV, simpler model, early stopping.
- Precision vs recall? — false-positive vs false-negative cost; F1 when both matter.
- ROC-AUC vs PR-AUC? — PR-AUC for imbalanced data.
- Bagging vs boosting? — parallel variance-cut vs sequential bias-cut.
Algorithms
- Linear vs logistic regression? — continuous output vs probability classifier.
- How does a decision tree split? — Gini/entropy gain; prune to control depth.
- Random Forest vs XGBoost? — robust baseline vs tuned SOTA on tables.
- K-Means how-to + K choice? — elbow/silhouette; scale features first.
- SVM intuition? — max-margin hyperplane; kernels for non-linearity.
Applied + MLOps
- Train/val/test split? — 70/15/15 or time-based for series; no leakage.
- Handle imbalance? — class weights, SMOTE, threshold tuning, PR metrics.
- Feature leakage example? — using future info (e.g., post-event columns).
- How do you deploy? — FastAPI + Docker + monitoring for drift.
Case Round Framework
Clarify metric → explore data → baseline → error analysis → iterate → trade-offs. Narrate aloud; interviewers score process. Practice pack: ML interview prep hub.
10 More Rapid-Fire Questions (With Answers)
- Signs of leakage? — Val accuracy ≈ 100% or features dated after the target event.
- When logistic over trees? — Small data, need probabilities/calibration, or linear signal.
- Scaling for KNN/SVM? — Mandatory — distances distort on raw scales; trees don't care.
- Curse of dimensionality? — Sparse space + overfit as features grow; fix with selection/PCA.
- Ensemble diversity? — Uncorrelated errors cancel — mix models + seeds + features.
- Choosing k in CV? — k=5 standard; k=10 for small data; time-series needs forward chaining.
- Early stopping patience? — Stop N epochs after val stops improving; restores best weights.
- Imbalance: which metric? — PR-AUC/F1 over accuracy; tune threshold to cost.
- Feature importance vs SHAP? — Importance ranks; SHAP explains direction + interactions per prediction.
- Your model degrades in prod — first 3 checks? — Drift in inputs, pipeline/schema change, label delay — in that order.
7-Day Revision Plan
- Day 1–2: definitions aloud (record yourself, 2-min each).
- Day 3–4: 1 project end-to-end narration.
- Day 5: full mock with a friend/mentor.
- Day 6: fix weak answers only.
- Day 7: rest + revise flashcards. Sleep beats cramming.
Annotated Mock Round (ML, 20 min)
- "Walk me through your churn model." → Baseline → features → validation → metric tied to money ("recall 0.81 at 5% FPR saves ₹X"). Business framing wins.
- "Val 0.95, test 0.72 — diagnose?" → "Leakage or time-split violation; check feature timestamps, re-split temporally." Shows debugging maturity.
- "Why XGBoost here, not deep learning?" → "8k tabular rows: trees win; DL needs 100x data + tuning budget." Judgment over hype.
- "Deploy in one week?" → "FastAPI + Docker, batch endpoint first, log inputs for drift." Shippable answer.
Frequently Asked Questions
- Hardest ML interview topic?
Evaluation + debugging (leakage, imbalance, drift) — beyond algorithm trivia. - Do freshers get ML interviews?
Yes at startups + analyst-to-ML ladders; 2 deployed projects mandatory. - Maths needed?
Linear algebra + probability intuition; derive log-loss once, then apply. - Best 60-day plan?
Month 1: algorithms + metrics. Month 2: projects + 5 mocks.
Read also: Learn ML · ML Projects · DS Questions