Explain PCA and L2 Normalization in Machine Learning
Overview
This question evaluates proficiency with preprocessing techniques (PCA and L2 normalization), variance reduction and multicollinearity, baseline model selection (logistic regression versus alternatives), backpropagation and gradient computation, knowledge-informed machine learning, and decision-threshold tuning for ROC-style evaluation.
Technical ML Phone Screen: Preprocessing, Models, Backprop, and Thresholding
Context
You built a binary classifier and used a preprocessing pipeline that included PCA and L2 normalization before training. You evaluated models with ROC-type metrics and considered logistic regression alongside alternative baselines. The interviewer wants you to connect preprocessing to variance reduction, explain gradients/backprop in neural nets, and discuss threshold selection and evaluation trade-offs.
Questions
-
Preprocessing
- Why did you apply PCA and L2 normalization before training? In what order? How does this relate to variance reduction and multicollinearity?
-
Modeling choices
- Walk through logistic regression as your baseline and compare it to reasonable alternatives you considered (pros/cons, when each shines).
-
Backpropagation
- Describe backpropagation in modern neural networks at a high level, including how loss gradients are computed and propagated.
-
Knowledge-informed ML and thresholding
- What is knowledge-informed (knowledge-guided) machine learning? Give concrete ways you used domain knowledge.
- How did you tune the decision threshold, and under what conditions can (or cannot) both FPR and TPR improve? Describe your process and trade-offs.
Solution
Login required