Data Science interview questions and answers

Heterogeneous treatment effects are different lifts for different users, such as new versus power users.

Pre-declare a primary metric, a small guardrail set, and either a fixed horizon or a sequential rule.

Read only needed columns, parse dates explicitly, downcast numerics, and convert low-cardinality strings to categorical.

agg reduces groups to summaries, transform returns aligned same-shape columns, and filter keeps or drops entire groups based on a group-level condition.

The warning appears when pandas is unsure whether a slice is a view or a copy, so chained assignment like df[mask]['col'] = x may not persist.

A view shares memory with the base array, so in-place edits alias

Those operations can change strides without moving data, producing Fortran-order or non-contiguous layouts.

numpy.where builds a new array from a condition and two value arrays, which is clear for vectorized if-then-else.

The frame clause defines which rows in the partition participate in the window aggregate relative to the current row.

Reduce scanned bytes with partition pruning, pre-aggregate to the metric grain in an upstream table, and avoid SELECT * in joins.

Build an acquisition cohort key per user, join later activity events whose dates fall in defined periods, and aggregate distinct actives over eligible cohort size.

Use SUM(metric) OVER (PARTITION BY entity ORDER BY date ROWS UNBOUNDED PRECEDING) or an equivalent frame.

Proxies such as post-treatment customer-support tags, future-dated snapshots, or IDs assigned only after conversion can encode the label.

Points can be typical marginally but rare jointly, which a pair of histograms will miss.

Distribution shift is a change in the input, label, or joint distribution after you ship.

MNAR means the missing value itself influences whether it is observed, such as users hiding high debt.

Multiple imputation draws several completed datasets from a model of the missing values, analyzes each, and pools estimates so uncertainty from missingness is reflected.

If missingness is MAR given covariates, you can model the probability of being observed and weight complete cases by the inverse of that probability.

Robust regressions downweight large residuals through a loss such as Huber, keeping all points but limiting their influence on coefficients.

Join logs, device data, and business rules: extra zeros, duplicate beacons, and timezone jumps look like errors, while consistent high spend across months looks real.