Data Science interview questions and answers

Percentile caps computed with test rows included use future or holdout extremes to define the training transformation.

Independent scales let you slide two series until they look perfectly correlated, which is visual rhetoric rather than evidence.

Starting bars above zero exaggerates small differences and violates the expectation that bar length maps to magnitude.

A DAG encodes assumed causal directions so you can find back-door paths that must be blocked and colliders that must not be conditioned on.

Difference-in-differences compares the before-after change in a treated group with the before-after change in a control group.

Propensity scores estimate the probability of treatment given observed covariates

An instrument is a variable that affects the treatment but does not affect the outcome except through treatment, and is not otherwise confounded with the outcome.

Importance sampling estimates expectations under a target distribution by drawing from a different proposal and reweighting with density ratios.

Weights correct for unequal sampling probabilities, nonresponse, and calibration to known population totals.

When you sample a large fraction of a finite population without replacement, the usual infinite-population variance formulas overstate uncertainty.

If variance is infinite or tails are extremely heavy, the usual sqrt(n) normality of the mean can fail or need enormous n.

A mixture is a weighted combination of component distributions, such as casual users and power users mixed in one histogram.

A Q-Q plot compares sample quantiles with theoretical quantiles of a reference distribution, often normal.

Autocorrelation is correlation of a series with its own lags.

Differencing subtracts lagged values to remove a stochastic trend and move toward stationarity, as in ARIMA's integrated part.

Granger causality tests whether lagged values of X improve prediction of Y beyond Y's own lags.

Time-series CV uses expanding or sliding windows so training always precedes validation, preserving order.

A North Star is the current best single measure of delivered customer value, such as weekly retained collaborative documents.

LTV estimates discounted future profit from a customer

If you reward a metric, teams will optimize the number rather than the underlying value, which is Goodhart's law.