Case Study

Explain why the learning curve is jagged for a skewed classifier and choose an appropriate subset strategy.

Case context: A team is training a model to detect machine failures from sensor logs. Only 10% of the training records correspond to failures. To build a learning curve, the team draws random training subsets of sizes 20, 50, 100, and 500 from the full training set. The resulting curve is very uneven, especially for the smallest subsets.

Question: Why is the curve so noisy at small subset sizes, and what sampling change should the team make? Under your proposed change, how many failure examples should a subset of size 20 contain?

Sample answer: The curve is noisy because random sampling from a dataset with only 10% failure cases causes large variation in the class mix for small subsets. With only 20 examples, the subset might accidentally contain too few or too many failures, which changes training behavior a lot from one draw to another. The team should build class-balanced subsets that preserve the original class proportions. For a subset of 20, 10% should be failures, so it should contain exactly 2 failure examples and 18 normal examples.

Key points:

  • Random small subsets from imbalanced data can have unstable class counts.
  • That instability makes the learning curve jagged.
  • Use subsets that match the original class proportions.
  • A subset of 20 should contain 2 failure examples at a 10% rate.

Rubric: The user must explain the noise as a consequence of high variance in class composition from random sampling, recommend matched-proportion subsets, and correctly calculate that 10% of 20 is 2.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related