Data Science interview questions and answers

Supervised learning uses labelled data to predict outcomes, such as classification and regression. Unsupervised learning finds patterns in unlabelled data, such as clustering.

Overfitting is when a model learns noise in training data and performs poorly on new data. It can be reduced with cross-validation, regularisation, simpler models, more data and early stopping.

Bias is error from overly simple assumptions; variance is error from sensitivity to training data. The goal is to balance both to minimise total error on unseen data.

Options include removing rows or columns, imputing with mean, median or mode, using model-based imputation, or adding an indicator for missingness — depending on the data and the amount missing.

A confusion matrix summarises classification results as true positives, true negatives, false positives and false negatives, and is used to calculate precision, recall and accuracy.