In a nutshell, the approach is kind of similar to the Netflix Challenge in that it's an ensemble of a bunch of different algorithms. However, while a good argument could be made that the Netflix winners were potentially overtraining on their "test" set by fitting so many different parameters, I think here there is clear out-of-band generalization.
1 comment
[ 92.0 ms ] story [ 586 ms ] threadHere's a good video that explains how this works: http://www.youtube.com/watch?v=3G2H3DZ8rNc
And a PDF version that goes into more detail: https://www-950.ibm.com/events/wwe/grp/grp004.nsf/vLookupPDF...
And the IBM Watson website: http://www.ibmwatson.com
In a nutshell, the approach is kind of similar to the Netflix Challenge in that it's an ensemble of a bunch of different algorithms. However, while a good argument could be made that the Netflix winners were potentially overtraining on their "test" set by fitting so many different parameters, I think here there is clear out-of-band generalization.