R. Sundararajan and A.K. Pal (India)
Reject option, Pattern recognition, Graded multiclass prob lems
We consider the problem of learning with the option to re ject in graded multiclass problems (GMP), i.e., problems where there exists a gradation among the class labels. We examine two aspects that distinguish GMP from standard multiclass problems: a) the performance metric of inter est in GMP may not be strict accuracy; accuracy within a bandwidth of the desired class may be permissible in some situations, and b) the ambiguity expressed by the classifier may simply reflect the underlying ambiguity in the class la beling itself. To deal with these issues, we extend the exist ing rejection schemes for multiclass problems to cover the GMP case. Firstly, we take the existing rejection schemes and calibrate them to optimize measures other than strict accuracy. Secondly, we present an extension to the well known dual threshold scheme to deal with GMP. We illus trate these ideas using a credit rating prediction problem.
Important Links:
Go Back
