Case Study

How a product team should split effort after checking error sources

Situation: A recommendation model was evaluated on a development set, and the team separated the mistakes by source. They found that 88% of the errors came from the way the ranking score was computed, while 12% came from the search procedure used to choose the best output. The group has time for only one major improvement.

Question: Which part should the team work on first, and why? If they put most of their effort into the search procedure instead, what is the largest total error reduction they should expect from that choice?

Sample answer: The team should focus first on improving the ranking score calculation, because it accounts for 88% of the observed mistakes. If they spend their effort on the search procedure, the best realistic improvement is about 12% of the total error, since that is the portion caused by search. The higher-impact path is to fix the scoring side.

Key points:

  • First priority should be improving the ranking score calculation
  • 88% of the errors are traced to the scoring step
  • Search-procedure changes can at most remove about 12% of the errors
  • The decision should follow the measured breakdown of error sources

Rubric: Full credit requires recommending improvement of score estimation, citing the 88% figure, and correctly stating the 12% ceiling on search-based improvements.

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