← Meta Interview Insights

Meta·Machine Learning Engineer·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

Went through a Meta interview that touched on ML fundamentals. Just the one question from what I can tell, but it's the kind of thing that sounds basic until you actually have to explain it clearly under pressure.

Questions Asked (1)

Q1

Why does feature scaling and normalization matter, and when would you apply it?

Technical Trade-offsAlgorithms & Data Structures
Author's notes

I knew the answer but fumbled the delivery.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining feature scaling and normalization, then explain why they matter for model convergence and performance. Discuss when to apply them based on algorithm sensitivity and data characteristics, and mention trade-offs and practical considerations.

Pro tip: Emphasize that scaling is not always necessary and can even hurt performance for tree-based models; showing this nuance demonstrates deeper understanding and practical experience.

1. Define the concepts

Clearly distinguish between feature scaling (e.g., min-max normalization, standardization) and normalization (e.g., L2 normalization). Explain that scaling changes the range, while normalization changes the distribution.

2. Explain why it matters

Discuss how features with different scales can dominate distance calculations and gradient updates, leading to slow convergence or poor performance. Mention that many algorithms assume features are on similar scales.

3. Identify when to apply

Specify algorithms that require scaling (e.g., SVM, k-NN, neural networks, PCA) versus those that don't (e.g., decision trees, random forests). Also consider data characteristics like outliers and sparsity.

4. Discuss trade-offs and best practices

Mention that scaling can be sensitive to outliers, and that normalization may not be suitable for sparse data. Highlight the importance of fitting scalers on training data only to avoid data leakage.

5. Conclude with practical impact

Summarize that proper scaling can lead to faster training, better model performance, and more reliable results, but it's not a one-size-fits-all solution.

Key Points to Mention

  • Gradient descent convergence: scaling speeds up convergence by making the loss surface more spherical.
  • Distance-based algorithms: k-NN, SVM, and clustering rely on distance metrics and are sensitive to feature scales.
  • Tree-based models: decision trees and their ensembles are invariant to monotonic transformations, so scaling is unnecessary.
  • Regularization: scaling ensures penalties are applied uniformly across features.
  • Data leakage: always fit scalers on training data and apply to validation/test data.
  • Outliers: min-max scaling is sensitive to outliers, while standardization is more robust.

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