Description of my Physics Problem Approach
==========================================

This work was a last-minute solution with little testing and parameter 
tuning. 
As preprocessing, attributes with less than 20 distinct values were binarized,
constant attributes removed and attributes with a very high range (values 
>10.000) were logarithmized. I used the kernel logistic regression algorithm 
(<http://www-ai.cs.uni-dortmund.de/SOFTWARE/MYKLR/>) with a radial basis 
kernel, combined with feature selection (training on a 5% sample and 
optimising the CXE on another 5% sample). The final prediction was generated 
by training with the optimal features on all examples. The same prediction 
was used for all performance measures (threshold 0.5 for accuracy).

Stefan Rueping
Dortmund University, CS Department, LS 8
stefan.rueping@uni-dortmund.de