Explain normalization, regularization, CTR, imbalance handling
Company: Microsoft
Role: Software Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
You are interviewing for an applied ML role. Answer the following fundamentals clearly and concretely (you may use equations and practical examples):
1) **Layer Normalization vs. Batch Normalization**
- What are the key differences in how they compute statistics?
- How do they behave during training vs. inference?
- When would you prefer one over the other (e.g., CV vs. NLP/LLMs, small batch sizes, RNNs/Transformers)?
2) **Regularization terms/techniques**
- Compare common regularization approaches such as **L2 / weight decay**, **L1**, **dropout**, **early stopping**, and **data augmentation**.
- What model behaviors do they encourage, and what are typical pitfalls?
3) **CTR (Click-Through Rate) prediction**
- Outline a practical approach to CTR prediction: data/feature setup (categorical + continuous), model families (e.g., logistic regression, factorization machines, deep models), training objective, and evaluation.
- Mention online concerns such as calibration and serving constraints.
4) **Handling imbalanced data**
- Describe strategies to train and evaluate models when positives are rare.
- Include both data-level and algorithm-level techniques, and appropriate metrics.
Quick Answer: This question evaluates mastery of normalization methods, regularization techniques, click-through-rate modeling, and class-imbalance strategies, with emphasis on model training and inference behavior, generalization, feature engineering, evaluation metrics, and production considerations.