inconsistency removal should depend on weighting #6
Description
Currently, weighted majority voting is done per class and as a separate step from inconsistency removal. Thus, the following situation can occur:
- GoodModel predicts class A, makes no statement about B (weight 100)
- BadModel predicts class B, makes no statement about A (weight 1)
- Ensemble predicts both A and B since all available models (per class) agreed
- Inconsistency removal comes in and, since A and B are disjoint, only one should be predicted. It decides this at random since it does not know how the ensemble decided originally
Goal
- Inconsistency removal decides based on the summed weights for disjointness (and trusts the one with the higher sum)
- For subsumption, this might be more tricky as correcting in different directions might introduce new inconsistencies