Amazon Data Scientist Interview Questions 2026

Tip: Classification predicts a discrete category (spam/not spam). Regression predicts a continuous value (house price). Logistic regression is classification despite the name — a common exam trap.

Tip: Bias: error from wrong assumptions (underfitting — model too simple). Variance: error from sensitivity to training data (overfitting — model too complex). Goal: sweet spot that generalises. Regularisation trades some variance for lower bias.

Tip: k-Fold CV splits data into k folds, trains k times each using a different fold as validation. Averages performance across folds for a more reliable estimate than a single split. Especially important for small datasets.

Tip: L1 (sum of absolute weights): produces sparse models by driving some weights to exactly 0 — acts as feature selection. L2 (sum of squared weights): shrinks all weights towards 0 but rarely to exactly 0. Use L1 for feature selection, L2 for general regularisation.

Tip: Precision = TP/(TP+FP) — of all predicted positives, how many are correct. Recall = TP/(TP+FN) — of all actual positives, how many did we catch. F1 = harmonic mean. High-precision when false positives are costly; high-recall when false negatives are costly.

Tip: Gradient descent: update weights opposite to the gradient of loss. Batch: uses ALL training data per step — accurate but slow. SGD: one sample per step — noisy but fast. Mini-batch: k samples per step — best of both, what deep learning uses (k typically 32–256).

Tip: Overfitting: model memorises training data, poor on unseen data. Detect: training accuracy significantly exceeds validation accuracy. Prevent: cross-validation, regularisation, dropout (neural nets), early stopping, more training data, feature reduction.

Tip: STAR format. Quantify impact wherever possible: "reduced churn by 8%", "improved conversion by 12L/month." The more specific the number, the more credible the answer.

Tip: Class imbalance: if 95% of data is class A, a model that always predicts A gets 95% accuracy without learning anything. Check precision/recall for the minority class. Also check if the metric the client cares about is accuracy at all — it often is not.

Tip: Feature engineering: creating new input features from existing data. Examples: log-transforming skewed salary data, extracting day-of-week from a timestamp, creating interaction terms. Good features often matter more than algorithm choice.

Tip: Lead with the business recommendation, not the methodology. Use plain language: "the model correctly identifies 9 out of 10 at-risk customers" beats "recall is 0.9". Use visuals. Avoid model-jargon entirely unless asked.

Tip: Python: general-purpose, better ML libraries (sklearn, PyTorch), production-deployment friendly. R: superior for statistical analysis, publication-quality plotting (ggplot2), bioinformatics. In Indian tech companies, Python is overwhelmingly preferred.