Opened 14 years ago

Last modified 13 years ago

#270 new discussion

Support for different methods to aggregate classifiers in EnsembleBuilder

Reported by: Markus Ringnér Owned by: Markus Ringnér
Priority: major Milestone: yat 0.x+
Component: classifier Version:
Keywords: Cc:

Description

In EnsembleBuilder? it is hard-coded that the prediction for each committe member is averaged to form the ensemble prediction. The supervised classifiers typically return continuous values for their predictions and these values are averaged. Alternatively each classifier could first be rounded to 0/1 votes for each class and the ensemble prediction would then be the majority vote. For example bagging was originally bootstrapping of samples to generate classifiers and the classifiers were aggregated by an unweighted majority vote.

We need some kind of structure to accomodate different aggregation rules. See also ticket:92.

Change History (1)

comment:1 Changed 13 years ago by Markus Ringnér

This is also related to deciding the winning class based on the continuous outputs from a classifier output. As it is now in KNN/NCC/..., decisions about winning classes are not implemented and left to the user. We should also keep in mind that for NCC the winning class has the smallest output (a distance to a centroid) whereas KNN and other have a large number for a winning class (e.g. the number of nearest neighbors in the class).

We should provide a general utility for this which could be used by all classes? Design ideas? EnsembleBuilder could then make use of this as part of one aggregation strategy.

Note: See TracTickets for help on using tickets.