Confusion matrix interface suggestions
Opened this issue · 2 comments
neubig commented
The new "confusion matrix" feature is a great start, but I felt the interface is a bit hard to understand:
- "F1 by confusion matrix" is not correct. It should just be "confusion matrix".
- And also, usually a "confusion matrix" will have, for each true tag
true_tag
and predicted tagpredicted_tag
:cooccurrence(true_tag, predicted_tag)/count(true_tag)
, but right now it's justcooccurrence(true_tag, predicted_tag)/total_count
. So we should either change the naming or the calculation.
cc @PaulCCCCCCH
PaulCCCCCCH commented
Thanks for creating the issue.
- Currently we display
<metric> by <feature description>
as the title. I guess we can simply remove the metric. - For the accuracy, I can fix the calculation is another PR.
By the way, the axes seem weird. Are these the real feature values for the task, or is there a bug?