|Title||Alternative OVA Proposals for Cooperative Competitive RBFN Design in Classification Tasks|
|Publication Type||Conference Paper|
|Year of Publication||2013|
|Authors||Charte, Francisco, Rivera Antonio J., Pérez-Godoy M.D., and del Jesus M. J.|
|Conference Name||12th International Work-Conference on Artificial Neural Networks (IWANN 2013)|
|Conference Location||Tenerife (Spain)|
In the Machine Learning field when the multi-class classification problem is addressed, one possibility is to transform the data set in binary data sets using techniques such as One-Versus-All. One classifier must be trained for each binary data set and their outputs combined in order to obtain the final predicted class. The determination of the strategy used to combine the output of the binary classifiers is an interesting research area. In this paper different OVA strategies are developed and tested using as base classifier a cooperative-competitive RBFN design algorithm, CO2RBFN. One advantage of the obtained models is that they obtain as output for a given class a continuous value proportional to its level of confidence. Concretely three OVA strategies have been tested: the classical one, one based on the difference among outputs and another one based in a voting scheme, that has obtained the best results.
Alternative OVA Proposals for Cooperative Competitive RBFN Design in Classification Tasks