Pinterest ML Interview: Model Evaluation Metrics Guide

Overview

This question focuses on model evaluation fundamentals you’ll be asked about in Pinterest ML foundation rounds. You’ll need to explain how to estimate model performance reliably (train-test split vs. cross-validation), choose appropriate evaluation metrics for different tasks, and reason about the bias–variance trade-off in both theory and practice.

Core content

You should be able to compare k-fold cross-validation variants (stratified, leave-one-out) and explain when to prefer each. Discuss metrics (accuracy, precision, recall, F1, ROC-AUC, PR-AUC) and their applicability for imbalanced classification and ranking problems. Show practical diagnostics (train/validation loss curves, learning curves) and strategies to mitigate overfitting (regularization, early stopping, simpler models) or underfitting (feature engineering, increasing model capacity).

Flow you might be asked to follow

1. Define how you would split data and justify the choice (time series, stratification).
2. Pick metrics for a concrete scenario and explain trade-offs (precision vs recall).
3. Diagnose bias vs variance using curves and validation scores.
4. Propose fixes and discuss impact on metrics.

Skill signals

Interviewers look for: solid understanding of cross-validation, correct metric selection for imbalanced data, ability to derive and interpret bias–variance decomposition, and actionable remediation steps (regularization, ensembling, calibration). Be ready to discuss unsupervised evaluation challenges and practical examples from projects.