Build and evaluate bad-link classifier
Company: Google
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
You have 1,000 URLs labeled as bad or good and a much larger unlabeled pool, with bad links rare. Design features and train a logistic regression. Explain your evaluation plan under class imbalance: stratified K-folds, ROC-AUC vs PR-AUC, calibration (reliability curves), and why accuracy is misleading. Choose a decision threshold by minimizing expected misclassification cost given asymmetric costs. Discuss class weighting or resampling, leakage checks, monitoring for dataset shift between labeled and production traffic, and an offline-to-online validation plan with shadow or canary deployment.
Overview: This question evaluates proficiency in applied machine learning classification, including feature design, training a logistic regression, handling severe class imbalance, selecting evaluation metrics and calibration, choosing thresholds under asymmetric costs, and planning offline-to-online validation and monitoring.
Read the full Google Data Scientist interview experience this question came from