Explainers Supplementary Material

Hamlet-Ness

For this example, we train Explainers to identify Shakespeare's Hamlet, amongst all of his plays, effectively building a measure of "Hamlet-ness."

To make things a little more challenging, we consider the 36 plays by act (so there's 180 total acts). For training the Explainer, there are 5 positive examples (the 5 acts of Hamlet) and 175 negatives.

The diagram uses color encoding to score the 180 different acts, for 5 different explainers. The scale is a purple/green diverging scale (so dark purple is very "non-hamlet", white is ambivalent, and deep green is very "hamlety".

The SVG files work in any browser with good SVG support. However, the HTML pages with embedded SVG seem to appear wrong in browsers other than Chrome. For some reason, the text gets messed up (probably a CSS issue).

The data used to generate the figures is available here.

When you look at this, notice how the first explainer uses 3 variables and is able to identify the first 4 acts of Hamlet - although some other plays have acts that are more Hamlety than Hamlet. Very different Explainers can pick out the last act. This might imply that the 5th act of Hamlet is quite different than the first 4.

The rightmost 2 explainers try to built regular linear SVMs that can pick out the acts of Hamlet from the pack. As you can see, getting good performance requires quite a complex SVM.