I found that regardless of the regularization term or features I used at the default
STEP_SIZE and SCALE_STEP, my performance peaked at around .40 .
It would thus appear that the saturation point of particular framework is tied just as
closely to the sampling parameters as it is to the underlying features.
After I set STEP_SIZE to 2 and SCALE_STEP to 1.3, my performance jumped from around .4 to
around .7 .
