LibSVM problem

Username

Hi,

I used the GridSearchParameterOptimization operator and the ProcessLog operator to find the best C value. The result was C=75 (accuracy: 76%). If I try to setup a process without GridSearchParameterOptimization and with the C value of the LibSVM learner set to 75 (via GUI or ParameterSet) I get an accuracy of only 73%. I can repeat these processes as often as I want, the accurays are always the same. The accuracys of C=75 is always higher if the GridSearchParameterOptimization operator was used. How can this be possible?

Thanks

TobiasMalbrecht

Hi,

I assume you used a cross validation inside the parameter optimization. The difference in accuracies probably occurs, because the example set partitions in both cross validations differ which happens when you use the global random generator for both cross validations. You can force the cross validation to sample identically by using the same local random generator. This can be done by setting the parameter [tt]local_random_seed[/tt] of the [tt]XValidation[/tt] operator to an identical positive value (e.g. 100).

Hope this was helpful,
Tobias