Model Evaluation
Testing a model to understand how well it performs and where it fails.
Which parcel labels does the reader miss?
A fictional parcel reader gets 95 of 100 destination codes right on held-out labels. A closer look changes the next test.
Clear print · 90 labels
The reader gets all 90 right. These labels dominate the overall figure.
Faded print · 10 labels
It reads 5 correctly and misses 5. Accuracy in this group is 50%.
The overall 95% and the faded-label 50% answer different questions. Preserve the group sizes so readers can interpret both.
Choose the next test from the error
If faded labels are common on one route, collect an independent sample from that route before relying on the reader there. Record the sampling dates, label conditions, and manual reference readings. A second test can then show whether a change actually helped the affected group.
Follow the idea
Precision helps when evaluation centers on the quality of alerts.
Model monitoring checks what happens on later shipments.
Related concepts (7)
Examples in the AI guides
More chapter examples (7)
Related Yenra articles (5)
Source · scikit-learn’s guide to scoring predictions
Return to Model Evaluation in the A–Z glossary