Explain core ML fundamentals
Overview
This question evaluates a candidate's mastery of core Machine Learning fundamentals including the bias–variance trade-off and regularization, gradient derivation for logistic regression with L2 regularization, evaluation metrics (ROC-AUC vs PR-AUC), cross-validation and data leakage, class imbalance mitigation techniques, model selection between tree-based and linear approaches, and calibration. It is commonly asked because it probes both conceptual understanding and practical application in the Machine Learning domain, testing theoretical reasoning about trade-offs and metrics as well as the ability to apply validation, evaluation, and model-selection principles in realistic settings.
ML Fundamentals — Onsite Interview Task
Context: Answer the following fundamentals as if in an onsite ML Engineer interview. Assume binary classification unless noted. For logistic regression with L2 regularization, use y ∈ {0,1}, feature matrix X ∈ R^{N×d}, parameters (w, b), sigmoid σ(z) = 1/(1+e^{-z}), and do not regularize the bias.
- Explain the bias–variance trade-off and how regularization techniques help.
- Derive the gradient for logistic regression with L2 regularization.
- Compare ROC-AUC versus PR-AUC and state when each is preferable.
- How do you detect and prevent data leakage in cross-validation?
- What methods handle class imbalance (e.g., weighting, resampling, thresholding) and how do they affect calibration?
- When would you prefer tree-based models over linear models, and why?
- How do you assess calibration and improve it?
Solution
Login required