Update: What this 95% figure means is that when they randomly select one white and one black from the sample and ask the model which is which it get the races right 95% of the time, compared to the 50% it would get from just guessing.
“Fig. 2 shows the prediction accuracy of dichotomous variables expressed in terms of the area under the receiver-operating characteristic curve (AUC), which is equivalent to the probability of correctly classifying two randomly selected users one from each class (e.g., male and female).“
The left column above is stuff black people like a lot more than white people do, and the right column is stuff white people like a lot more than black people do (small towns, Davie Bowie, Harley Davidsons, being nice to animals, sterility, etc.).
As a study of Twitter showed in 2010, black people really like being black.
Some of the best discriminator likes for whether a Facebook user drinks alcohol sound like a Norm McDonald Weekend Update from 1996:
I Like Watching Raindrops Silently Race Across My Window and Cheering for Them
Meeting Someone Who Is Also Drunk and Immediately Becoming Best Friends.
And then there's gay v. straight males:
Yet, another thing that could be researched from this database is how common are exceptions to stereotypes. For example, I had a gay friend who was a huge college football and major league baseball fan. Obviously, that's unusual, but I can't really estimate quantitatively how unusual. I'd say "an order of magnitude," but that's pretty much what I say for every quantitative estimate for which I don't say "oh, about fifty-fifty."
In this study, the model of Facebook likes could get gay man vs. straight right 88% of the time, but lesbian vs. straight right only 75% of the time. (Here's my article "Why Lesbians Aren't Gay" from 19 years ago, which lists about three dozen likes and dislikes that tend to distinguish gays from lesbians.)