PMML preprocessing ports


For me it is not clear what is the purpose of the PMML preprocessing port (for example in k-Means, Fuzzy c-Means, RProp MLP Learner, Decision Tree Learner, SVM Learner nodes). Is it there to support the kernel methods? (In case yes, is there a reason for for example Distance Matrix Calculate (for k-Medoids) not having this option? I guess with the R to PMML node it would be possible to generate the proper functions, but I had no luck with it.*) Or are these better suited for ensemble learning? (This pdf suggests it is quite general, and suitable for ensemble, but I am curious about other options too. Maybe it requires a not-yet-implemented node with the common kernel functions for kernel methods? Or is there a node for that? Maybe even a looping too for adjusting the parameters of the kernel method?)

Thanks, gabor

*: in the attached workflow: Execution of R script failed: Calls: toString -> pmml

Hi gabor, I guess this publication answers all of your questions:

Ah, of course this explains only the first part. I don't think kernel methods can be incorporated, but you can build a workflow an inject models generated with R at any point. The exception indicated that your model could not be translated into PMML. Can you please check back with the corresponding R method.

Thank you, I guess the usual kernel methods are not easy to achieve or this way it might be not possible. The R expression was quite simple, but maybe not well formed for PMML:

R<-1 / (1 + exp(-R$"a1"))