Preview — Pro guide
You are seeing a portion of this guide. Sign in and upgrade to unlock the full article, quizzes, and interview answers.
Sections
Related Guides
Bias-Variance Tradeoff & ML Debugging
Machine Learning
How to Approach an ML Interview Round at FAANG
Machine Learning
Cross-Validation Strategies: K-Fold, Time Series, Nested CV, and Leakage-Proof Pipelines
Machine Learning
Imbalanced Classification: Metrics, Class Weights, SMOTE, and Threshold Tuning
Machine Learning
Loss Functions: Choosing the Right Objective for Every ML Problem
Machine Learning
How to Structure ML Interview Answers — Derivations and Debugging
Execution playbook for FAANG ML interview rounds. Covers the model selection framework, canonical whiteboard derivations (bias-variance, softmax cross-entropy gradient, backprop through an MLP), the model debugging playbook, and practical patterns for imbalanced classification, calibration, and cross-validation.
Answering ML Questions Is a Mechanical Skill
Strong ML candidates do not invent answers from scratch in 45 minutes. They run a playbook: a sequence of frameworks for the recurring question types — model selection, derivation, debugging, and tradeoff. The skill is not raw cleverness; it is knowing which framework applies and assembling the answer under time pressure.
This page is the mechanical playbook for ML interview answers. It assumes you already know how to approach the round (covered on the companion page) and focuses on what to say and why.
The rule that organizes everything: match the answer's structure to the question's structure. "Tell me about overfitting" is a multi-layer question (definition + signal + fix menu) and needs a multi-layer answer. "Derive backprop" is a math question and needs structured algebraic work. "Your model is overfitting in production" is a debug question and needs ranked hypotheses with named diagnostics. Mismatching the structure is the most common L4-cap error.
The second rule: every answer ends with a connection to a real model or production decision. After deriving softmax cross-entropy gradient (p - y), finish with: "this is why label smoothing helps — the one-hot target makes the gradient saturate at confidence 1, which is what causes the modern-classifier overconfidence Szegedy et al. (2016) documented." The connection signal is what separates rote knowledge from practical understanding.
The Model Selection Framework — Run This for Every 'Pick a Model' Question
Step 1 — Identify the data shape (30 seconds)
Tabular (rows of features), sequence (time-series, text), image, graph, or multimodal? Data shape eliminates 80% of model classes. Tabular -> GBDT (LightGBM, XGBoost) is the production standard. Sequence -> Transformer or LSTM. Image -> CNN (ResNet, EfficientNet) or Vision Transformer. Graph -> GNN (GCN, GraphSAGE). Mismatch is the most common architectural mistake.
Step 2 — Identify the loss function (30 seconds)
Regression -> MSE (L2, sensitive to outliers) or MAE (L1, robust, non-smooth) or Huber (smooth + robust). Binary classification -> binary cross-entropy (logistic loss); for class imbalance, focal loss (Lin et al., 2017, gamma=2 typical). Multi-class -> categorical cross-entropy with softmax. Ranking -> pairwise (RankNet) or listwise (ListNet, LambdaRank). Mismatch causes either no signal or wrong signal.
Step 3 — Identify the regularization (30 seconds)
L2 weight decay (lambda=1e-4 typical for neural nets, 1.0 typical for ridge regression on standardized features) — uniform shrinkage. L1 — sparse solutions, feature selection. Dropout (0.1-0.5) for fully-connected and attention layers. Early stopping for neural nets. For trees: max_depth (6 default for XGBoost), min_samples_leaf (10-50 typical), subsample (0.8 typical for stochastic gradient boosting).
Step 4 — Identify the evaluation metric (30 seconds)
Imbalanced classification -> AUC-ROC for ranking quality, AUC-PR for performance on the rare class, F1 for fixed-threshold tradeoff. Multi-class -> macro-F1 (equal weight per class) or weighted-F1 (weighted by support). Ranking -> NDCG, MAP, MRR. Regression -> RMSE (sensitive to outliers, what most papers report) or MAE (median-friendly). Choose the one that aligns with the business objective.
Step 5 — Identify the validation strategy (30 seconds)
IID data + large -> 80/20 split + stratified for classification. Small data -> k-fold cross-validation (k=5 or 10 typical). Time-series -> chronological split or expanding-window CV; never random split. Grouped data (multiple samples per user) -> grouped k-fold to prevent leakage. Imbalanced -> stratified sampling by class label.
Step 6 — Justify with one sentence per choice
For each of the 5 above, name the alternative you considered and rejected. 'I picked LightGBM over XGBoost because the data has high-cardinality categoricals and 5M+ rows; LightGBM has native categorical handling and leaf-wise growth converges faster at this scale (Ke et al., 2017). XGBoost is the safe alternative if I needed broader ecosystem support.'