When the submission period for the General Assembly and Michelin Guide data science challenge ended last week, there were 25 total hopefuls. Twelve were students in the GA data science immersive course (it was a required project), but the remaining 13 were from outside — real D.C. data scientists in the real world.
This was cool, instructor Joseph Nelson said, because it allowed the students (then in their 11th week of the course) to see how their own skills measure up. It wasn’t just another project, it was a “validation.”
And the GA students did pretty well — three of the top five place holders were students. But ultimately it was Brendan Sudol, a developer at 18F, who won out.
Sudol subsequently told GA his approach was “Nate Silver-esque.”
“I really don’t know too much about the food scene in D.C., so I relied on averaging and aggregating other reviews,” he said. Sudol submitted 14 restaurants, and when the guide was released on Oct. 13 it became clear that he’d managed to correctly identify seven of them as well as the number of stars each of those seven got. Not bad.
See Sudol’s work here:
So what’d even the data scientists get wrong?
Nelson told Technical.ly the most missed restaurant that made it into the guide was the Inn at Little Washington, though that’s arguably more Michelin’s fault than anything (the Guide set the expectation ahead of time that all restaurants featured would be in D.C., but then included this at the last moment). Most data work also suggested that Komi and Rasika would be among the restaurants to win stars, and they did not. Interestingly, what the data scientists got wrong is quite similar to what the qualitative, food critic predictors got wrong as well.
Nelson said that the response to the project both from within the class and from the outside participants was positive enough that he now hopes to do a similar kind of challenge with each cohort.