I'm using VW 7.4 to do some binary classification:
cat train | vw -k --binary --cache_file cache -f model --compressed --passes 10 --loss_function hinge
At test time, I can access the scores using:
cat test | vw -t -i model -p pred
What I don't get is that, if I add --bootstrap 4 to training, then the resulting pred file contains 3 columns. I don't know which of them corresponds to the scores of the prediction.
When you use --bootstrap <N> vowpal-wabbit adds a confidence [min max] interval resulting from the multiple bootstrapping samples, to each prediction line:
prediction tag min max
The first two fields remain as they were without --bootstrap:
prediction tag
(tag is optional; it is echoed from the respective example in the data-set, and its use is recommended)
You can find more details in Zhen Qin's presentation slides on the new features he contributed (including but not limited to --bootstrap)
Unrelated, but has to be mentioned:
You don't have to feed vowpal-wabbit from stdin using cat file |. You can save a process by passing the data-set file directly to vw:
vw ... filename