advanced

ML Foundations

Probability and statistics, generalization and validation, and working with class imbalance — the reasoning layer under every model.

ml-foundations

Practice: 11 questions →

Sections

Contents

Detailed explanation

ML Foundations

Before any model there is a layer of reasoning that decides whether its numbers mean anything: the probability that generates the data, the discipline that keeps a score honest, and the arithmetic of a rare positive class. A model can fit perfectly and still be worthless if the score you trust was inflated by leakage, or if the metric measures the majority class instead of the thing you care about. This topic is the foundation the rest of ML stands on.

The thread through all three layers is the same interview reflex: name what a number measures before you trust it.

Topic map

Probability & statistics — distributions and what they model, expectation and variance, Bayes and base rates, MLE vs MAP, and what a p-value and a confidence interval actually claim (and never claim).
Generalization — bias vs variance, overfitting vs underfitting, honest train/val/test splits, cross-validation schemes, regularization, learning curves, and the leakage that quietly inflates every score.
Imbalanced data — why accuracy lies on rare classes, precision/recall/F1, PR-AUC vs ROC-AUC, resampling and class weights and their risks, threshold selection, and calibration.

What interviewers probe

The common thread is disproving a confident misreading: a p-value is not the probability the null is true; a high validation score is not honest if a transform was fit before the split; 99.9% accuracy is not skill when the base rate is 0.1%. Get these three foundations right and every downstream model conversation gets easier.

Why it matters

Before any model there is a layer of reasoning that decides whether its numbers mean anything. Skip it and you will ship a model that scores 0.92 offline and 0.71 in production, quote a p-value as if it were the chance your hypothesis is wrong, or celebrate 99.9% accuracy on a problem where the answer is always "no". These are not exotic failures — they are the default outcome of trusting a metric you have not questioned, and interviewers probe exactly here to tell a practitioner from someone reciting formulas.

ML Foundations

The thread through all three layers is the same interview reflex: name what a number measures before you trust it.

Topic map

Probability & statistics — distributions and what they model, expectation and variance, Bayes and base rates, MLE vs MAP, and what a p-value and a confidence interval actually claim (and never claim).

Generalization — bias vs variance, overfitting vs underfitting, honest train/val/test splits, cross-validation schemes, regularization, learning curves, and the leakage that quietly inflates every score.

Imbalanced data — why accuracy lies on rare classes, precision/recall/F1, PR-AUC vs ROC-AUC, resampling and class weights and their risks, threshold selection, and calibration.

What interviewers probe