Interactive Confusion Matrices

Apple recently came out with a paper that had an interesting idea.

The paper describes possible improvements to a confusion matrix by allowing for nested classes as well as interactivity and I would argue that they some sensible ideas here.

The basic interface looks like this.

Pretty neat to add precision/recall/accuracy numbers there.

When you click to expand a class, you can see some more members.

That's pretty neat!.

You can keep an expanding for a deeper overview.

More depth.

You can play with a demo of this work here.

I like the direction they've considered here, although not every classification problem will have nested classes like those shown here. It's also a shame that the current implementation gives you a large dashboard which won't work very well when you're inside of a Jupyter notebook. But I do like the idea of exploring interactivity as a method to understand model performance better.