We wanted something like Netflix recommendations for research papers... because a nonzero number of us like reading those, and a certain ranty lecture from class stuck with us.
What it does
Presents the user with a paper title, author, conference, abstract, and URL to the paper. The user rates (on a scale of 1-3) how much they'd like to read the paper based on the abstract. A user's highly rated papers are noted and they can reference which papers (and respective URLs) they wanted to see.
How we built it
- Data Structure Sins™
Challenges we ran into
- Losing four hours at the beginning on uncooperative (read: ignorance) web development and a bad idea. Solution: Ditch shenanigans (read: Node.js) and get a better idea.
Accomplishments that I'm proud of
- "Justin, what are you proud of?", "We're not using Node.js. I don't know. Make one up."
- "Will, what are you proud of?", "...", "I'll come back to you."
- "Eric, what are you proud of?", "You can make one up for me, too."
- "Audrey, what are you proud of?", "..."
What we learned
- Sometimes you need Data Structure Sins™ when time is not a thing on your side.
- A bad idea turned into a joke idea that turned into this idea... four hours into the hackathon.
- In Node, half your code will comprise of import statements.
What's next for SparkPapers
- Transfer Learning: to generate synthetic data, which will serve as a better starting point.
- The use of ensemble methods; at the moment, we simply use DBSCAN.
- The full 36 GB of research papers and a V100.