Case Study

Sizing a dev set for a high-revenue product recommendation engine.

Case context: You are leading the machine learning team for a major e-commerce platform's product recommendation engine. The current model is already highly optimized. An engineer proposes testing a new algorithm that they estimate will improve accuracy by just 0.01%, but they note that a dev set of 10,000 examples won't be able to confirm this.

Question: Based on Ng's principles, should you invest in evaluating this tiny improvement, and what must you do to the dev set to test it properly?

Sample answer: Yes, the improvement should be evaluated because for mature, high-value applications like product recommendations, a 0.01% gain directly and significantly increases company profits. To test it, you must increase the dev set size to much larger than 10,000 examples to reliably detect such a small change.

Key points:

  • 0.01% improvements are highly valuable in mature applications.
  • Small improvements directly affect company profits.
  • A dev set larger than 10,000 examples is needed to detect small improvements.

Rubric: The student must recognize that evaluating the 0.01% improvement is worthwhile due to its financial impact in mature applications, and conclude that the dev set size must be increased well beyond 10,000 examples.

0

1

Updated 2026-06-13

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Machine Learning Strategy

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Yearning @ DeepLearning.AI