class |
PredictionErrorIQR |
Post-processor that removes outliers using a coarse IQR approach on the predictions errors of one or more classifiers.
parameters:
- list of classifiers
- foreach classifier; number or folds, an IQR multiplier, number of iterations or number of consecutive non-removal iterations before stop.
algorithm:
foreach classifier
loop
do xval, get predictions, remove all examples where error > percentile75+IQR*multiplier of errors
stop if done num_iterations, or consecutive zero removals
|