We both wanted to sharpen our data science skills and make a super meta hackathon project at the same time.

What it does

Given only text from a devpost project description, our app predicts whether or not that project will win an award.

How we built it

First, we scraped every single project on devpost. (Python, beautiful-soup) We then trained classifiers on "bag-of-words" representations of the projects. (scikit-learn: PCA, SVM) Finally, we deployed our model to the web to fulfill it's destiny as an automated hackathon judge.

Challenges we ran into

  1. Scraping devpost was a pain in the rear. Why do they not have an official API???
  2. Text data can be counterintuitive. We did a bunch of stuff we though should substantially improve model performance. Our test accuracy decreased. FML frownie-face.
  3. It turns out, our model only predicts losses. If 30 percent of all projects get some kind of prize, we can achieve 70 percent accuracy by classifying all projects as non-winners. We have failed to show that word counts can distinguish winning hackathon projects from the rest.

Accomplishments that we're proud of

  • We scraped almost 50 MB of project descriptions from devpost
  • We learned some tough lessons about what can go wrong when building a classifier.
  • We made it to the final round in the cup stacking competition.

What we learned

  • How to build a classifier and put it on the web.
  • "Better than a coin-toss" doesn't mean your model isn't worthless.

What's next for Will It Win?

Deep Learning and hackathon project creation using "Deep Dream." trippy...

Share this project: