9 Data Science interview questions and answers

The model learns noise in the training data instead of the pattern, so it scores well on training data and badly on new data.

Detect it by comparing training and validation scores, using cross-validation or learning curves. Fix it with more data, a simpler model, regularisation, early stopping or fewer features.

High bias means the model is too simple and underfits. High variance means it is too sensitive to the training data and overfits.

Expected error is roughly bias² + variance + noise that no model can remove. Choose the model complexity that gives the lowest validation error.

Precision = TP / (TP + FP): of the cases you flagged, how many were right. Recall = TP / (TP + FN): of the real cases, how many you caught.

Favour recall when missing a case is costly, such as disease screening or fraud. Favour precision when false alarms are costly, such as a spam filter. F1 is their harmonic mean; the decision threshold moves the balance.

If 99% of rows are negative, a model that always predicts negative scores 99% accuracy while catching nothing.

Use the confusion matrix, precision, recall, F1 and PR-AUC. Handle the imbalance with class weights, resampling and a tuned decision threshold.

Leakage is when the model sees information during training that it won't have at prediction time: future data, features derived from the target, or scaling fitted on the full dataset.

In k-fold cross-validation, fit all preprocessing inside each fold, for example with a scikit-learn Pipeline. For time series, split by time instead of shuffling.

L1 (lasso) adds the sum of absolute weights to the loss and pushes some weights to exactly zero, so it also selects features.

L2 (ridge) adds the sum of squared weights and shrinks all weights smoothly, which handles correlated features well. Elastic net combines both.

A random forest trains many deep trees independently on bootstrap samples and random feature subsets, then averages them. It mainly reduces variance and works well with little tuning.

Gradient boosting trains shallow trees one after another, each fitting the errors of the previous ones. It mainly reduces bias, is often more accurate, and needs careful tuning of learning rate and number of trees to avoid overfitting.

Define one primary metric and the hypothesis. Calculate the sample size from the baseline rate, the minimum effect worth detecting, power (usually 80%) and significance level (usually 5%).

Randomise by user, run the full planned duration, and check that the split between groups matches the plan. Analyse with a suitable test, such as a two-proportion z-test, and report the effect with a confidence interval. Don't stop early because results look good, unless you planned a sequential test.

First rule out pipeline bugs: missing features, changed units, schema changes.

Then compare live feature distributions with training data to find data drift, and check whether the relationship between features and labels has changed (concept drift). Break errors down by segment, retrain on recent data if needed, and add monitoring and alerts so you catch it sooner next time.