Machine Learning interview questions

How to organize customer clustering for evaluating questions that a specific customer has no historical data for?

What is the difference between gradient boosting and Random Forest?

How many parameters are in a fully connected layer with N inputs and K outputs (with bias)?

ROC-AUC = 0.9. Duplicated each positive object 7 times, negative — 4 times. How will ROC-AUC change?

Moderation: two thresholds — automatic rejection and sending for manual review. What's more important for each threshold: precision or recall?

Explain decision trees, splitting criteria, Random Forest vs Gradient Boosting. How do they affect bias-variance?

What parameter-efficient fine-tuning (PEFT) methods for large models do you know?

What classification metrics do you know? What are their differences?

Explain precision and recall. In which tasks is what more important?

What neural network architectures have you used in your work?

How to start building a trading strategy for bitcoin, having price, volume and orderbook data?

What AI/ML trends will dominate in the next 5 years?

What approaches to parallel neural network training do you know?

Explain the Self-Attention mechanism in transformers

Why is gradient boosting called 'gradient'? What does it optimize?

What approaches were used for molecular energy optimization? What are the model inputs and outputs?

How do categorical features work in boosting models? What encoding options exist?

Why CatBoost? What makes it unique compared to other libraries?

Explain bias-variance decomposition for Random Forest and gradient boosting.

Why does the transformer use Layer Normalization instead of Batch Normalization?