← Adobe Interview InsightsMachine Learning Engineer
I started with MCAR/MAR/MNAR which felt right, but I could tell they wanted me to actually connect that diagnosis to the downstream choice, not just recite definitions.
Structure your answer as a systematic pipeline: first diagnose the missingness mechanism (MCAR, MAR, MNAR), then assess the impact on your model and business context, and finally select an imputation strategy that balances statistical validity with practical constraints. Emphasize that the best approach depends on the data type, missingness pattern, and downstream model, and that you validate choices through experimentation.
Pro tip: Always start by quantifying the missingness and its relationship to the target variable—if missingness itself is predictive, consider adding a missing indicator feature rather than just imputing. Also, be prepared to discuss how you'd handle missing data in production, ensuring your imputation logic is reproducible and monitored.
Determine whether data is missing completely at random (MCAR), at random (MAR), or not at random (MNAR) by analyzing patterns and correlations with other variables. Use statistical tests (e.g., Little's MCAR test) and visualizations to guide your understanding.
Evaluate how much data is missing, which features are affected, and how missingness might bias your model or business outcomes. Define what a successful imputation looks like in terms of model performance and interpretability.
Select from simple (mean, median, mode), model-based (KNN, MICE), or advanced (deep learning) methods based on data type, missingness mechanism, and computational budget. Consider adding missing indicators if missingness is informative.
Test the imputation strategy using cross-validation and compare model performance against baselines (e.g., dropping missing rows). Monitor for overfitting and ensure the imputation doesn't leak information from the test set.
Deploy the imputation as part of a reproducible pipeline, ensuring it handles new data consistently. Set up monitoring to detect shifts in missingness patterns and retrain or adjust as needed.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.