Data science interviews in India test three things: whether you understand the statistics under the models, whether you can evaluate a model honestly, and whether you know what happens after a model is trained. Every question below has the answer an interviewer is listening for, and where a question carries a trap, the trap is named.<\/p>\n
These check that your understanding goes past the library call.<\/p>\n
What is the difference between data science, data analytics and machine learning?<\/strong><\/p>\n Data analytics answers what happened and why, mostly with querying, aggregation and descriptive statistics. Data science includes that and extends into predicting what will happen, using statistical models and machine learning. Machine learning is the subset concerned with algorithms that learn patterns from data rather than following rules a person wrote.<\/p>\n What is the difference between supervised and unsupervised learning?<\/strong><\/p>\n Supervised learning trains on data where the answer is known, so the model maps inputs to a labelled output, as in predicting whether a customer will churn. Unsupervised learning has no labels and finds structure instead, as in clustering customers into segments. Semi-supervised and reinforcement learning sit outside both, and are worth naming briefly.<\/p>\n What is the difference between classification and regression?<\/strong><\/p>\n Both are supervised; the difference is the output type. Classification predicts a category, such as whether a loan will default. Regression predicts a continuous number, such as how much a customer will spend next quarter. Predicting a probability of default is still classification, because the decision taken from it is categorical.<\/p>\n What is overfitting and how do you spot it?<\/strong><\/p>\n Overfitting is a model learning the noise in the training data rather than the pattern, so it performs well on data it has seen and poorly on data it has not. You spot it from a large gap between training and validation scores. The usual remedies are more data, simpler features, regularisation and early stopping.<\/p>\n What is the bias-variance trade-off?<\/strong><\/p>\n Bias is error from the model being too simple to capture the real relationship, and shows up as poor performance everywhere. Variance is error from the model being too sensitive to the training sample, and shows up as strong training performance with weak validation performance. Complexity moves you along the trade-off, and the goal is the point of lowest total error rather than either extreme.<\/p>\n Why do you need a validation set as well as a test set?<\/strong><\/p>\n The validation set is used to tune hyperparameters and choose between models, so information from it leaks into your choices. The test set is held back and touched once, giving an unbiased estimate of performance on unseen data. Tuning against the test set and then reporting that score is the commonest way candidates overstate a model, and interviewers do ask.<\/p>\n What is cross-validation and when do you use it?<\/strong><\/p>\n K-fold cross-validation splits the data into k parts, trains on k minus one and validates on the rest, rotating until every part has served as validation, then averages the scores. It gives a more stable estimate than a single split, which matters most on small datasets. Time series needs a forward-chaining split instead, because shuffling lets the model train on the future.<\/p>\n What is feature engineering and why does it matter more than the algorithm?<\/strong><\/p>\n Feature engineering is creating the input variables the model learns from: ratios, aggregates over a time window, date parts, encoded categories. In most business problems a well-engineered feature set with a straightforward model beats a sophisticated model on raw columns. Interviewers ask this to see whether you have solved a real problem or only run algorithms on clean datasets.<\/p>\n This is where most data science interviews are decided, and where the traps cluster.<\/p>\n Why is accuracy a poor metric for imbalanced data?<\/strong><\/p>\n Because a model that predicts the majority class every time scores well without learning anything. On a fraud dataset where 1% of transactions are fraudulent, always predicting legitimate gives 99% accuracy and catches nothing. Use precision, recall or precision-recall AUC instead, chosen by which error costs the business more.<\/p>\n Explain precision and recall in a business context.<\/strong><\/p>\n Precision is the share of your positive predictions that were correct, so it says how much you can trust an alert. Recall is the share of actual positives you caught, so it says how much you are missing. Raising the threshold pushes precision up and recall down, and where to sit is a business decision about the cost of a false alarm against a missed case.<\/p>\n What does ROC AUC actually measure?<\/strong><\/p>\n It measures how well the model ranks a randomly chosen positive above a randomly chosen negative, across every threshold. A value of 0.5 is no better than chance and 1.0 is perfect separation. On heavily imbalanced data it stays flatteringly high because the large negative class dominates the false-positive rate, which is why precision-recall AUC is the honest companion.<\/p>\n What is regularisation, and what is the difference between L1 and L2?<\/strong><\/p>\n Regularisation penalises the size of the model coefficients so it cannot fit noise by pushing weights to extremes. L1, or Lasso, penalises absolute values and can shrink coefficients exactly to zero, which effectively selects features. L2, or Ridge, penalises squared values and shrinks towards zero without eliminating, handling correlated predictors more gracefully.<\/p>\n How does a decision tree decide where to split, and what does a random forest add?<\/strong><\/p>\n A tree evaluates candidate splits and picks the one that most reduces impurity, measured by Gini or entropy for classification and variance for regression. A single deep tree overfits badly. A random forest trains many trees on bootstrapped samples with a random subset of features at each split, then averages them, cutting variance sharply.<\/p>\n How is gradient boosting different from a random forest?<\/strong><\/p>\n A random forest builds trees independently and averages them, which mainly reduces variance. Gradient boosting builds trees in sequence, each fitting the residual errors of what came before, which mainly reduces bias. Boosting usually wins on tabular business data but is more sensitive to hyperparameters and noisy labels, and overfits if you keep adding trees without early stopping.<\/p>\n What is data leakage and how does it show up?<\/strong><\/p>\n Leakage is information in your training features that would not exist at prediction time, or that is derived from the target. The signs are a validation score that looks too good and a model that collapses in production. Common causes are scaling or imputing before the split, including a field populated only after the outcome occurs, and joining an aggregate computed over the whole dataset including the future.<\/p>\n What is the curse of dimensionality, and what does PCA do about it?<\/strong><\/p>\n As the number of features grows, the data becomes sparse, distances mean less and the sample size needed to learn reliably grows fast. Principal component analysis reduces dimensions by projecting onto the directions carrying the most variance. The cost is interpretability, since components are combinations of your original features and nobody in a business meeting can reason about component three.<\/p>\n How do you handle imbalanced classes?<\/strong><\/p>\n Start with metrics that do not reward the majority class, then set the decision threshold deliberately rather than accepting 0.5. Class weights in the loss function are usually the cleanest lever, with undersampling the majority or synthetic oversampling of the minority as alternatives. Resample only the training folds, because oversampling before the split leaks near-duplicates into validation.<\/p>\n Choosing an evaluation metric is itself an interview question, and the metric follows from what a wrong answer costs. These are the ones that come up most, with the trap attached to each.<\/p>\nIntermediate level: modelling and evaluation<\/h2>\n