Explainers Supplementary Material

Figure 2: An American-ness Parallel Coordinates Plot

This figure is a parallel coordinates plot of 5 different explainers of "American-ness." It's a really simple parallel coordinates plot, and it also suffers from the fact that since the data is so highly quantized, the resulting explainers are quantized, and the lines pile on top of each other even worse than usual.

Here, the cities that were annotated as in the USA are marked blue, rather than gray. San Jose is marked in dark blue. The fact that it so consistently scores low in "American-ness" suggests that either Silicon Valley is quite different than the rest of the US, or that San Jose in the data is not San Jose, California. Maybe its actually San Jose, Guatamala.

These explainers were chosen to be American-ness explainers with 3 variables, small integer coefficients, and to be as diverse as possible while still having nth scores greater than .75.

If you're curious as to what the explainers themselves look like, here they are...

Tooltips over the blocks will show the city names. The boxplots give some indication that these are all reasonable explainers (the innter quartiles of US cities are above the inner-quartiles of the other cities). I find it interesting that by several explanations of American-ness, Johannesburg and Pretoria score higher than the US cities.

I should also note that the data has some variables where being high in the variable is good, and some where being high in the variable is bad. I didn't make the data (it comes from a contest on Buzzdata.com).

You might wonder how something with an MCC score (Mathews correlation coefficient) of 0 can possibly be a good explainer (in terms of correctness). Equivalently, how can one metric (nth) say that an explainer has good correctness, but another (mcc) says it doesn't. In this case, I think the problem is that the "optimization" to choose the threshold (required to get the best MCC score) is picking a bad threshold. Since the optimization maximizes accuracy (not mcc), sometimes it decides its better to always say no (if no is the more common class) even though this gives chance performance. A better answer would be to implement the threshold optimization to maximize mcc score (rather than simple accuracy).