Learn Before
Decide when relabeling bias is acceptable in two planned uses of a result.
Case context: A team is building a customer support classifier and first wants to use the result only for internal model debugging. Later, the same result may be included in a conference paper.
Question: Determine when the relabeling bias is acceptable and when it becomes a problem.
Sample answer: The bias is acceptable if the result is used only for internal debugging of the product. It becomes a problem once the team plans to present the result in a conference paper. It is also a problem if the team needs the result to serve as a fully unbiased estimate of test performance.
Key points:
- Internal debugging can tolerate the bias
- Conference-presentation use makes the bias problematic
- A requirement for an unbiased test-performance estimate also makes it problematic
- The judgment depends on how the result will be used
Rubric: The response must clearly distinguish internal development use from conference-paper use and classify the bias appropriately in each case. Mentioning the need for an unbiased performance estimate strengthens the answer.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
When is it acceptable to clean up noisy labels even if the labels are not perfectly objective?
True or False: Correcting labels before reporting results in a research paper can create problematic label-correction bias.
Label-correction bias may be acceptable when the goal is simply to build a _____ or service.
Match each scenario to whether label correction is acceptable.
Order the steps for deciding whether label-based correction is acceptable
Why Label Cleanup Can Be Acceptable in Some Settings
Decide when relabeling bias is acceptable in two planned uses of a result.
When is label correction a poor choice?
A project needs an unbiased held-out accuracy estimate. How should it view relabeling errors in that set?
A small product team and a publishable research study should always handle label fixes in the same way.