Subset Correction of a predicted result
This method restrict a multi-label learner to predict only label combinations whose existence is present in the (training) data. To this all labelsets that are predicted but are not found on training data is replaced by the most similar labelset.
An object of mlresult that contain the scores and bipartition values.
A matrix/data.frame with all labels values of the training dataset or a mldr train dataset.
A numeric value between 0 and 1 to use as base to
determine which values needs be reescaled to preserve the corrected
A logical value. If
If the most simillar is not unique, those label combinations with higher frequency in the training data are preferred. The Hamming loss distance is used to determine the difference between the labelsets.
A new mlresult where all results are present in the training labelsets.
The original paper describes a method to create only bipartitions
result, but we adapeted the method to change the scores. Based on the
base.threshold value the scores higher than the threshold value, but must be
lower are changed to respect this restriction. If
correction will be ignored.
Senge, R., Coz, J. J. del, & Hullermeier, E. (2013). Rectifying classifier chains for multi-label classification. In Workshop of Lernen, Wissen & Adaptivitat (LWA 2013) (pp. 162-169). Bamberg, Germany.