Data Science Interview Questions India 2026: ML, Python

Statistics is the foundation, interviewers test fundamentals ruthlessly:

1. Probability basics: P(A and B) = P(A) P(B) for independent events. P(A or B) = P(A) + P(B) - P(A and B). Conditional probability: P(A|B) = P(A and B) / P(B). Bayes theorem: P(A|B) = P(B|A) P(A) / P(B). Used in: spam classification, medical diagnosis, A/B test result interpretation.

2. Distributions: Normal distribution: symmetric bell curve; mean = median = mode; 68-95-99.7 rule (1, 2, 3 standard deviations). Central Limit Theorem: the sample mean of any distribution approaches normal as sample size increases (n ≥ 30 rule of thumb). Binomial: number of successes in n independent Bernoulli trials. Poisson: count of events in a fixed interval (customer arrivals, website errors). Understanding which distribution to model a business problem with is a common case question.

3. Hypothesis testing: Null hypothesis (H0): no effect (the new recommendation engine does not change CTR). Alternate hypothesis (H1): there is an effect. p-value: probability of observing the data (or more extreme data) if H0 is true. Significance level (α = 0.05): reject H0 if p-value < 0.05. Type I error (false positive): reject H0 when it is true. Type II error (false negative): fail to reject H0 when H1 is true. Statistical power: probability of correctly rejecting H0 when H1 is true. Common tests: t-test (compare means), chi-square test (categorical association), Mann-Whitney U (non-parametric mean comparison).

4. A/B testing: Randomly assign users to control (A) and treatment (B). Measure a primary metric (conversion rate, revenue per session). Ensure sample size is large enough (power analysis pre-experiment). Avoid peeking (checking results before reaching the target sample size inflates false positive rate). Multiple testing problem: if you test 20 metrics, one will be significant by chance, use Bonferroni correction or false discovery rate control.

ML algorithm questions at Indian data science interviews:

1. Linear regression: Models the relationship between a continuous outcome and one or more predictors. Assumes: linearity, independence of errors, homoscedasticity (constant variance of residuals), no multicollinearity. Coefficients estimated by minimising the sum of squared residuals (OLS). Regularisation: Ridge (L2) shrinks coefficients toward zero; Lasso (L1) can shrink some coefficients to exactly zero (feature selection); ElasticNet combines both.

2. Logistic regression: For binary classification. Models the log-odds of the outcome as a linear combination of predictors. Output: probability via sigmoid function. Threshold: typically 0.5 but tunable. Interpretable: coefficient exponentiated = odds ratio. Used as a strong baseline at Indian companies before trying tree-based or neural network models.

3. Decision trees, Random Forest, and Gradient Boosting: Decision tree: recursively splits data to reduce impurity (Gini or entropy for classification; MSE for regression). Prone to overfitting; limited depth or pruning required. Random Forest: ensemble of decision trees trained on bootstrap samples; each split considers a random subset of features. Reduces variance (overfitting) via averaging. XGBoost / LightGBM: gradient boosting; builds trees sequentially, each correcting the errors of the previous. Dominant model at Indian product companies for tabular data. LightGBM: faster than XGBoost; better on large datasets; leaf-wise tree growth.

4. Evaluation metrics: Classification: accuracy (misleading for class-imbalanced data), precision (of all predicted positives, how many are actually positive), recall (of all actual positives, how many did we predict), F1-score (harmonic mean of precision and recall), AUC-ROC (probability that the model ranks a random positive higher than a random negative). Regression: RMSE (penalises large errors), MAE (robust to outliers), MAPE (percentage error; misleading when actuals are near zero), R-squared (proportion of variance explained).

5. Handling class imbalance: Fraud detection, churn prediction: the minority class (fraud, churn) is often < 1% of data. Oversampling minority: SMOTE (Synthetic Minority Over-sampling Technique). Undersampling majority: random undersampling. Class weights: most sklearn models accept class_weight='balanced'. Threshold tuning: lower the classification threshold to catch more positive cases (improves recall at the cost of precision).

Python coding and case study skills:

1. Pandas and NumPy: DataFrame operations: readcsv, head, info, describe, valuecounts, groupby + agg, merge, pivottable, apply. Handling missing data: isnull, fillna, dropna. String operations: str.contains, str.split, str.replace. Date parsing: pd.todatetime, dt.year, dt.month, dt.dayofweek. NumPy: vectorised operations, broadcasting, array slicing. Interview tasks: given a DataFrame of transactions, compute monthly revenue per user; find users who purchased in January but not February; compute rolling 7-day sum.

2. Scikit-learn pipeline: Preprocessing: StandardScaler (mean 0, std 1: required for algorithms that use distance), MinMaxScaler, OneHotEncoder (for categorical features), SimpleImputer (fill missing values). Pipeline: chain transformers and estimator: Pipeline([('scaler', StandardScaler()), ('clf', LogisticRegression())]). Cross-validation: crossvalscore(model, X, y, cv=5, scoring='rocauc'). Grid search: GridSearchCV(model, paramgrid, cv=5).

3. Feature engineering: The biggest impact on model performance comes from feature engineering, not model choice. Numerical: log transform (right-skewed), polynomial features, interaction terms, binning. Categorical: one-hot encoding (low cardinality), target encoding (high cardinality: replace category with mean of target). Date: day of week, hour of day, days since last event, is_weekend. Text: TF-IDF, word embeddings (Word2Vec), pre-trained BERT features.

4. Business case study pattern: Common case types at Indian companies: (a) Metrics decline: 'Daily orders on Swiggy fell 15% yesterday. Walk me through your investigation.' Clarify: is it affecting all cities, all order types, specific payment methods, specific device types? (b) Product launch: 'Design an experiment to test a new recommendation algorithm.' Define success metric (CTR, conversion), experiment unit (user, session), sample size, duration, guardrail metrics. (c) Model build: 'Build a model to predict churn.' Clarify: what is churn? (30 days inactive) What data is available? Feature engineering → model selection → evaluation → deployment → monitoring.