Opened 15 years ago

Last modified 15 years ago

#96 closed enhancement

Feature selection should not be part of SupervisedClassifier — at Version 4

Reported by: Peter Owned by: Peter
Priority: critical Milestone: yat 0.2
Component: classifier Version:
Keywords: Cc:

Description (last modified by Peter)

needs #108 #109 #110

The simple reason is think modules. The longer reason is that present design is not general enough. I have the following user case:

I wanna build an ensemble of classifiers to cross-validate the perfromance. In this procedure, I wanna select feature using a ranked list (InputRanker?), though I do not want to select the top features in the ranked list, but rather say features in position 151-251.

There is no way I can do that as it is now. The feature selection in SupervisedClassifier? could of course be extended, but I think it would become to complex and it is better to lift out the whole thing to a new class. Polymorphism could be really useful here, besides having the obvious class be constructed from a Score, start feature and end feature, one could imagine having a class constructed from a vector of indices. I suppose a sensible interface would be to only have one function get(const DataLookup2D& entire_data, DataLookup2D& selected data). Other variables needed are set in corrsponding constructor.

Change History (4)

comment:1 Changed 15 years ago by Peter

Status: newassigned

comment:2 Changed 15 years ago by Peter

Description: modified (diff)

comment:3 Changed 15 years ago by Peter

Description: modified (diff)

comment:4 Changed 15 years ago by Peter

Description: modified (diff)
Note: See TracTickets for help on using tickets.