Machine Learning Meetup Notes:2011-4-13
Anthony Goldbloom from Kaggle Visits
- Link to his talk: PPT presentation
- Guy used random forests to win HIV competition. Word "random forests" is trademarked. Dude taught himself machine learning from watching youtube videos. Random forests are pretty robust to new data.
- Used caret package in R to deal with random forests.
- Kaggle splits test dataset into two, uses half for leaderboard.
- Often score difference between winning model and second place is not statistically significant. So they award prizes to top few. Might impose restrictions on execution time of model.
- Performance bottoms out in competitions within a few weeks in general. This seems to be due to all the information being "squeezed" out of the dataset at that point.
- Chess rating competition: build a new rating system that more accurately produces the results. The performance still plateaued, but took longer.
- Most users of kaggle are from computer science and statistics, followed by economics, math, biostats.
- Tools people use:
- R: lots of american users
- Python: although it's lower on the list, people are successful with it
- R packages used: Caret, RFE, GLM, NNET, Forecast
- Heritage Prize
- Real shit is going down may 4th, with release of all datasets.
- Ends in 2 years. No rush.
- Four prizes in total, given out throughout the next two years.