Explainers Supplementary Material

Figure 4: Comedicness, 6 ways.

Figure 4 of the paper shows 6 different explainers for "comedic-ness". (This was introduced in Figure 0, which was not part of the paper. You might want to check Figure 0 so you can read this diagram).

Here there are 6 explainers, all of which are trained to put comedies above the non-comedies. The leftmost one uses only two variables, with unit weights. The next one uses three variables with unit weights, and still gets a few wrong.

The third from the left shows that you can get perfect correctness using only 3 variables, but it requires a more complicated combination of those variables - this explainer is very sensitive to small values in "MoveBody" (it has a very large weight for it).

While that while this explainer gets all of the answers right (all comedies are above all non-comedies), the "margin" between the lowest comedy and the highest non-comedy is quite small. (look at the wiskers on the modified boxplot). You can get a sense of the performance tradeoffs using the scagnostics plot (Figure 0s).

The theory behind Support Vector Machines suggests that maximizing this margin (relative to the total range of the output function) is desireable. The three explainers to the right illustrate this: they have successively larger margins (notice the huge gap between "A Comedy of Errors" and "Hamlet" in the rightmost one), at the expense of some very complex functions.