Professional Machine Learning Engineer

By Google · Question Mode

QuestionQ47

Scaling prototypes into ML models

You were asked to investigate production-line component failures using sensor readings. After receiving the dataset, you find that fewer than 1% of readings are positive examples representing failure incidents. You have attempted to train several classification models, but none converges. How should you address the class-imbalance issue?

Explanation

Downsampling the majority negative class increases the frequency of failure examples in training batches, which supports effective learning and faster convergence. Upweighting the downsampled majority-class examples corrects the artificial class-prior bias introduced by sampling, while a training sample with 10% positives provides a substantially more usable balance. Google Machine Learning Crash Course: Class-imbalanced datasets

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!