When Evaluation Scores and Product Needs Disagree
Case context: A team is building a model that flags abusive comments in an online forum. Their development metric ranks Model A above Model B. After manually inspecting outputs, the team decides Model B is better for the product because Model A still lets too many harmful comments slip through.
Question: Identify the sign that evaluation is off track, name the most likely underlying problem, and state what the team should change and what they should do right away after changing it.
Sample answer: The sign is that the development metric prefers Model A even though people reviewing real examples prefer Model B. The most likely problem is that the metric is rewarding the wrong behavior for the product, so it does not match what the team truly wants to improve. The team should revise the evaluation metric to reflect the real goal, then tell everyone on the project about the new metric and direction.
Key points:
- Notice the disagreement between the metric ranking and the product ranking.
- Conclude that the metric is optimizing the wrong objective.
- Update the metric so it matches the project goal.
- Communicate the change to the whole team.
Rubric: The response must identify the ranking disagreement as the sign of a problem, explain that the metric is targeting the wrong objective, recommend changing the metric, and state that the team should communicate the new direction to everyone involved.
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 Evaluation Data Does Not Match Deployment Data
When Repeated Validation Checks Distort Model Selection
When the Metric Rewards the Wrong Goal
When should your validation setup be revised?
True or False: If your initial validation split or evaluation metric turns out to be poorly chosen, you cannot revise it without abandoning the project.
If your evaluation metric no longer reflects your main objective, what should you change?
What is the clearest sign that your dev/test set or evaluation metric may need revision?
If a validation set or metric turns out to be poorly matched to the real goal, the team should rebuild the whole project before making any changes.
What to revise when the evaluation no longer matches the goal
Match each reason a validation metric can mislead the team to the recommended remedy.
What should a team do when its evaluation setup stops matching its goal?
When Validation Data Does Not Match Deployment Data
After revising your dev/test sets or evaluation metric, updating the project documentation is enough; the team does not need to be told about the new direction.
What should be expanded after repeated tuning to the validation set?
Match each situation to the underlying problem category it illustrates.
Order the reasoning steps for deciding whether to replace an evaluation metric that no longer matches the product goal.
When validation results stop matching the best product choice
When Evaluation Scores and Product Needs Disagree
What should a team do after the development set stops guiding decisions?