The request
“Create a confusion matrix heatmap for an email-intent classifier with actual rows and predicted columns Order, Return, Shipping and Account. Use counts by actual row: Order 91, 7, 3, 2; Return 9, 76, 5, 4; Shipping 2, 6, 84, 7; Account 3, 2, 8, 79.”
Then tryChange the Account predicted Shipping count to 5.Remove the visual scale from the chart.
What is in this drawing
Read the decisions behind it.
01Use customer intents
Align classes with real queues.
02Review related intents
Order and Return errors may need new examples.
03Track validation date
Compare the same test protocol over time.
Use when one incoming email has one agreed routing intent.
More like this
Other examples in this family.
FAQ
Common questions
Why include four intents?
It reflects distinct queues in a common support-routing workflow.
What does an off-diagonal count mean?
It is a message routed to a different intent from its actual label.
Should this run without review?
That depends on the impact of each routing error and the team's controls.