Google

An evaluation metric can be flawed even if it is computed correctly, because it may reward an outcome that is not the project’s real objective. When that happens, the score should not be used to compare models, and the team should redefine the metric so it matches the true goal.

When the Metric Rewards the Wrong Goal

In a pet-photo moderation app, one classifier can achieve higher overall accuracy than another and still be a poor choice if it sometimes approves violent images that should have been blocked. A model can score better on accuracy and still fail the product's real safety goal.

Accuracy Is Not Always the Right Metric

One way to improve a weak evaluation metric is to assign a very large penalty to a specific mistake that the metric currently treats too lightly, such as allowing unsafe content to pass through.

Using a Strong Penalty for a Critical Error Type

If the current evaluation metric can no longer be trusted, replace it with a new metric that clearly defines the team’s objective. Do not rely for an extended period on informal manual comparison among candidate models.

Set a New Evaluation Goal When the Current Metric Is Unreliable

What should a team do if its evaluation metric is rewarding the wrong outcome?

Choose the verb that best completes the statement.

A score that measures the wrong target should not be used to ____ the best model.

Match each metric concept to its description.

What should a team do after noticing its metric points to the wrong goal?

Question:
What happens if a machine learning evaluation score measures something different from the project’s real objective, and what should the team do to fix it?

Sample answer:
If the score is not aligned with the project’s true goal, it stops being a dependable way to compare models. The team may end up choosing a model that looks good on the score but performs poorly on the real task. The fix is to replace or redesign the evaluation metric so it matches the outcome the project actually needs.

Key points:
- The score is tracking the wrong target.
- It can no longer be relied on to pick the best model.
- The remedy is to change the evaluation metric so it reflects the real objective.

Rubric:
The response should clearly state that (1) the metric is misaligned with the project goal, (2) it is no longer trustworthy for model selection, and (3) the appropriate response is to revise the evaluation metric.

What to do when an evaluation score favors the wrong goal

Case context:
A payments company is comparing several fraud detection models. It selects the model with the highest overall accuracy on a validation set. In practice, the chosen model misses many fraudulent transactions because fraud is rare and the metric rewards mostly correct predictions on legitimate transactions.

Question:
Explain why the current selection process is failing and state what the team should change.

Sample answer:
The process is failing because overall accuracy does not reflect the project's real goal, which is to catch fraud effectively. A model can score well on accuracy while still missing too many bad transactions. The team should replace the current metric with one that matches the business objective, such as a metric that emphasizes fraud detection or the cost of missed fraud.

Key points:
- The metric measures the wrong thing for the business goal.
- High accuracy can hide poor fraud detection when the class is imbalanced.
- The team should use a metric aligned with the true objective, especially the cost of missed fraud.

Rubric:
The student must identify that the metric is misaligned with the project objective and recommend changing to a metric that better reflects fraud detection performance.

Diagnose a metric mismatch in a fraud detection model selection process.

Question:
If the score used for model selection rewards the wrong behavior, how should the team treat that score, and what is the next step?

Sample answer:
The score should no longer be treated as a dependable way to choose among models. The team should redefine the evaluation measure so it reflects the actual project goal.

Key points:
- The current score is not dependable for selecting the best model.
- The evaluation measure should be changed to match the real objective.

Rubric:
The answer should say that the metric is not trustworthy for model selection and that the team needs to change the evaluation metric.


Learn Before

Related