← Adobe Interview InsightsMachine Learning Engineer

Adobe·Machine Learning Engineer·Technical Phone Screen·Senior

Senior
Apr 2026

Summary

Adobe ML Engineer interview, technical phone screen focused entirely on ML fundamentals. One meaty question about data leakage that went deeper than I expected, covering like four different subtypes. Not a vibe check at all, they wanted real depth.

Questions Asked (1)

Q1

What is data leakage in machine learning? Walk through the different forms it can take, such as target leakage, train/test contamination, time-based leakage, and leakage introduced during preprocessing. How do you detect and prevent each type in practice?

Technical Trade-offsRoot Cause AnalysisAlgorithms & Data Structures
Author's notes

This question started simple and then just kept expanding.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Define data leakage as any information in training that wouldn't be available at prediction time, then systematically walk through each form with concrete examples and detection/prevention strategies. Emphasize that leakage often stems from improper data splitting or feature engineering and can be caught by rigorous validation and monitoring.

Pro tip: Always simulate production conditions by splitting data temporally or by entity, and use pipelines to encapsulate all preprocessing steps to avoid leakage. Mention that leakage can also occur in feature stores or online serving, so consistency between training and serving is key.

1. Define data leakage

Explain that data leakage occurs when a model uses information during training that will not be available at inference time, leading to overly optimistic performance and poor generalization.

2. Enumerate forms of leakage

Describe target leakage, train/test contamination, time-based leakage, and preprocessing leakage, giving a concrete example for each.

3. Detection techniques

For each form, outline how to detect it: e.g., unexpected feature importance, performance drop on holdout, temporal validation, and pipeline audits.

4. Prevention strategies

Detail prevention: use pipelines, split data before preprocessing, respect temporal order, avoid using future data, and validate with proper cross-validation.

5. Real-world practices

Discuss how to maintain consistency between training and serving, monitor for leakage in production, and conduct code reviews to catch subtle issues.

Key Points to Mention

  • Target leakage: using features that are proxies for the target or derived from it (e.g., 'days since last purchase' when predicting churn).
  • Train/test contamination: preprocessing (e.g., scaling, imputation) fit on full data before splitting, or duplicate rows across splits.
  • Time-based leakage: using future data to predict past events; need temporal splits and lag features.
  • Preprocessing leakage: fitting transformers on training+test data; use pipelines and fit only on training folds.
  • Detection: monitor for suspiciously high performance, check feature importance, use adversarial validation, and compare validation vs. test performance.
  • Prevention: use scikit-learn pipelines, split data first, use time-series cross-validation, and ensure feature availability at inference.

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