The final day of EDM was a bit more quiet than the previous day, which is more than acceptable because there was a lot of amazing ideas going around that day. John Stamper had a good keynote talk in regards to how the KDD 2010 Cup went when using the Educational Data from DataShop. It turns out that the National University of Taiwan has an entire course dedicated to winning the KDD Cup every year, as a result, they won this past year as well. Sadly there were very few people from the EDM community that participated, so on that point it was a bit of a disappointment. Luckily however, he did announce that there will be an EDM competition for generating these types of models. Rumor has it he is looking to find a substantial prize money award as well, so hopefully there will be more participation in the years to come.
Less is More
There was a cool paper about improving the speed and prediction power of knowledge tracing by using less data, in the work it was essentially discovered that when you use the entire set of past attempts, you don’t improve prediction any more than when you use just the past five attempts. The reason for this is more or less due to over-fitting. When you use all the past attempts the students who make more errors have more past attempts and that hurts the model. By using just last five or so actions you get a better average of the whole, because good students may get the correct answer in just one or two attempts. By using all the attempts the model indirectly favors students that do poorly. It was a good paper, and the discussion offered a fair amount of conclusiveness and closure, all in all less is more.
Q-Matrices
Lastly, Michel Desmarais presented a really cool work on deriving a Q-Matrix from data with Non-negative matrix factorization. This paper won the best paper award and rightly so, I really enjoyed this work and Dr. Desmarais did an excellent job presenting it, with a very thoughtful and easy to understand talk. I am a big fan of Q-Matrices as well as automatic methods for deriving knowledge components.
The End
Well that is it for EDM 2011, I had a wonderful time, Eindhoven was a great venue and there were a number of great works, cool research and wonderful contributions. I really look forward to next year’s conference and I hope everyone else is too. This is a pretty exciting field and just because the conference is over, stay tuned for other posts here about how EDM, the community and its related areas, grow and evolve over the course of the year. I hope you’ve enjoyed my reports and I look forward to seeing you all next year, until then happy data mining.
