Adobe·Machine Learning Engineer·Technical Phone Screen·Senior

SeniorPrefer not to say

Jul 2026Remote

Summary

Adobe ML Engineer screen that basically turned into a deep dive on data quality. One main technical question but it had a lot of branches and I wasn't fully prepared for how far they'd push it.

Questions Asked (1)

Walk me through how you handle missing data in a machine learning pipeline, from identifying why it's missing all the way through choosing an imputation strategy.

Technical Trade-offsData ModelingAlgorithms & Data Structures

Author's notes

Started okay by bringing up MCAR vs MAR vs MNAR, which seemed to land well.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer as a systematic diagnostic and decision process: first classify missingness mechanisms (MCAR, MAR, MNAR) and assess missingness patterns, then evaluate imputation methods against data type, missingness rate, and downstream model requirements. Emphasize that the best strategy depends on the trade-off between simplicity, bias, and computational cost, and always validate imputation impact via cross-validation.

Pro tip: Mention that you track missingness as a feature (e.g., missing indicator) when missingness itself is predictive, and that you prefer to impute inside cross-validation folds to avoid data leakage—this shows production maturity.

1. Diagnose missingness

Quantify missing rates per feature and per sample, and visualize patterns (e.g., missingno matrix) to distinguish MCAR, MAR, and MNAR. Interviewers want to see that you don't jump to imputation before understanding the mechanism.

2. Assess impact and constraints

Determine how much data would be lost by dropping rows/columns, check for leakage risks, and consider computational and latency constraints of the production pipeline. This frames imputation as a trade-off, not a default.

3. Choose imputation strategy

Select methods based on data type and missingness mechanism: simple statistics (mean/median/mode) for MCAR/MAR with low missingness, model-based (KNN, MICE, iterative imputer) for MAR with complex relationships, and domain-specific defaults or missing indicators for MNAR. Justify why you reject alternatives.

4. Implement and validate

Fit imputers only on training folds and apply to validation/test to prevent leakage; compare models with and without imputation using cross-validation and appropriate metrics. Highlight that imputation is part of the model selection process.

5. Monitor and iterate

In production, monitor missingness drift and imputation performance over time, and be ready to retrain or adjust the strategy as data distributions change. This shows end-to-end ownership.

Key Points to Mention

MCAR, MAR, MNAR definitions and how they influence imputation choice
Trade-offs between deletion, simple imputation, and model-based imputation (bias vs. variance vs. complexity)
Use of missing indicator features when missingness is informative
Avoiding data leakage by fitting imputers within cross-validation folds
Evaluation of imputation impact on downstream model performance (not just imputation accuracy)
Handling missing data in production: monitoring, drift detection, and fallback strategies

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.