Google

Precision is the proportion of examples the model predicted as cats on the dev or test set that were actually cats.

Precision for a Cat Detector

In many classification problems, improving precision can make recall worse, and improving recall can lower precision, so there is often a tradeoff between the two.

Precision and Recall Usually Compete

What does precision mean for a spam filter?

True or False: Precision is determined only from the examples a classifier predicts as positive.

Match each term to its description related to precision.

Order the steps to calculate precision for an image classifier.

Question:
Explain, using the definition of precision, why precision by itself may give an incomplete picture of a model that detects damaged packages in a shipping center.

Sample answer:
Precision measures the fraction of packages the model marked as damaged that were truly damaged. A model can achieve a high precision score by being very selective and only flagging the most obvious damage cases. But that does not tell us how many damaged packages the model missed. If many damaged packages are not flagged, the model may still be unsuitable even with strong precision. That is why precision is often considered together with recall.

Key points:
- Precision is the fraction of predicted-damaged packages that are actually damaged
- Precision does not count damaged packages the model failed to flag
- A model can raise precision by making fewer positive predictions
- A complementary metric such as recall is needed to see how many true damaged packages were found

Rubric:
Full credit requires defining precision correctly, explaining that it only evaluates the predicted-positive cases, and noting that it ignores missed damaged packages, which is why a second metric such as recall is needed.

Explain why precision alone can be misleading for a model that flags damaged packages.

Case context:
A model flags 80 emails as spam on a validation set. When the team checks those 80 predictions against the true labels, 68 really are spam and 12 are legitimate messages.

Question:
Using the definition of Precision, what is the Precision of this model on the validation set, and what does that tell you about its spam predictions?

Sample answer:
Precision is the share of predicted spam messages that are truly spam, so the Precision is 68/80 = 0.85, or 85%. This means that when the model marks an email as spam, it is right 85% of the time. It describes how dependable the spam alerts are, not how many spam emails the model found overall.

Key points:
- Precision = true positive predictions / all positive predictions
- 68/80 = 0.85 = 85%
- Higher Precision means predicted positives are more reliable
- Precision does not measure how many real spam emails were missed

Rubric:
Full credit requires correctly computing Precision as 68/80 and correctly explaining that it measures the reliability of the model’s positive predictions, not its coverage of all actual spam emails.

Compute a classifier’s precision on a validation set.

Question:
In one to three sentences, identify what set of items belongs in the denominator when calculating precision for a spam classifier.

Sample answer:
The denominator is the total number of messages the classifier predicted as spam, whether those predictions were correct or not.

Key points:
- Denominator is messages predicted as spam by the classifier
- Not all messages being evaluated
- Includes both true positives and false positives

Rubric:
Full credit for identifying the denominator as all messages predicted as spam by the classifier, not all messages in the dataset.

Learn Before

Related