Opened 15 years ago

Closed 15 years ago

#86 closed enhancement (wontfix)

SVM should use a MatrixLookup (and not KernelLookup) when feature selection is used.

Reported by: Markus Ringnér Owned by: Peter
Priority: major Milestone: yat 0.3 (Public release)
Component: classifier Version:
Keywords: Cc:

Description

NCC and SVM should implement the interface specified in Supervised classifier similarily so that they can be used changed smoothly when feature selection or ensembles are used. SVM should also have additional functionality working with Kernels internally but only when feature selection is not used.

Change History (3)

comment:1 Changed 15 years ago by Peter

There are a few functions noted with a @todo remove in the Kernel classes. Remove these when redesigning the MatrixLookup? and they should not be needed anymore.

comment:2 Changed 15 years ago by Peter

Status: newassigned

comment:3 Changed 15 years ago by Peter

Resolution: wontfix
Status: assignedclosed

Letting the SVM taking MatrixLookup? means the SVM must know how to create a KernelLookup?. This includes two choices, type of KernelFunction? and type of Kernel (Kernel_SEV or Kernel_MEV). My gut feeling is that is preferable to keep those underlying structes hidden to the SVM, meaning kernel methods work on KernelLookups? also when feature selection is included. A nice thing with the present design is that the Kernel completely takes care of the question whether analysis is weighted or not.

Note: See TracTickets for help on using tickets.