Pretty standard but I still fumbled the structure a bit.
Start by diagnosing why the data is missing and how much is missing, then choose a handling strategy based on the missingness mechanism and the downstream impact. Emphasize that the right approach depends on the business context and the modeling goal, and always validate that your choice doesn't introduce bias.
Pro tip: At Meta, interviewers care about scale and product impact—mention that you'd quantify the cost of dropping vs. imputing (e.g., lost users, biased metrics) and consider whether missingness itself is a signal (e.g., a user not clicking could be meaningful).
Identify which columns have missing values, the percentage missing, and the pattern (random vs. systematic). Determine if missingness is MCAR, MAR, or MNAR.
Clarify the analysis or model objective and how missing data could bias results. Consider whether missingness itself carries information (e.g., absence of an event).
Select from deletion (listwise/pairwise), imputation (mean/median/mode, regression, multiple imputation, ML-based), or flagging missingness as a feature. Justify trade-offs.
Test the impact of your choice via cross-validation, sensitivity analysis, or A/B testing if possible. Monitor for introduced bias or degraded performance.
Clearly document assumptions and methods, and communicate limitations to stakeholders. Ensure reproducibility and transparency.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.