Data Science interview questions and answers

axis specifies which dimension you collapse: axis=0 reduces over rows for a 2D array, leaving columns.

Vectorized ufuncs run in compiled code over contiguous memory and can use SIMD.

A boolean mask of the same shape selects positions where the mask is True, returning a one-dimensional collection of those values for a 1D array.

LEFT JOIN keeps every row from the left table and fills unmatched right columns with NULL.

Window functions compute aggregates or ranks over a partition while keeping row-level detail, using an OVER clause with partition and order keys.

A common table expression is a named subquery introduced with WITH, which you can reference like a temporary table.

DISTINCT removes duplicate full rows but does not compute sums, and it can hide that you joined explosively and then deduped.

NULL means unknown, so NULL = NULL is not true and WHERE col = NULL filters nothing

A correlation matrix screens pairwise linear associations among numeric features and can flag redundant predictors.

Leakage shows up as features that would not be known at prediction time, perfect separation, or timestamps after the label.

Duplicates inflate counts, double-weight users in means, and can be real repeated events or pipeline bugs.

Cardinality is the number of distinct values.

MCAR means missingness is unrelated to observed and unobserved values.

Median imputation is more robust on skewed numeric columns because a few giants do not dominate the fill value.

The fact of missingness can itself predict the outcome, for example users who skip billing fields.

Listwise deletion drops a whole row if any analysis variable is missing.

Values below Q1 minus one-and-a-half IQRs or above Q3 plus one-and-a-half IQRs are conventional box-plot outliers.

A common rule flags |z| greater than 3 under an approximate normal model.

Keep them when they are valid rare events you care about, such as whale spenders, outages, or fraud.

Winsorization caps extreme values at chosen percentiles instead of dropping rows, so sample size stays the same.