search

jpmml / r2pmml

R library for converting R models to PMML
GNU Affero General Public License v3.0
73 stars 18 forks source link

Add support for one class-SVM #17

Closed gilad19 closed 7 years ago

vruusmann commented 7 years ago

Can you clarify, which R package/function interests you? Is it kernlab::ksvm(.., type = "one-svm"), e1071::svm(.., type = "one-classification"), or something completely different?

gilad19 commented 7 years ago

Hi,

Thanks. I actually interested in both (kernlab::ksvm and e1071::svm). I guess if I was asked to choose I would pick kernlab::ksvm but it doesn't really matter (by the way, if I am not wrong the type of ksvm is "one-svc").

Thank again!

vruusmann commented 7 years ago

Decided to implement the e1071::svm() function first - it's based on a popular LibSVM implementation, and it supports model specification via the "formula" interface.

The r2pmml package encodes one-class SVM similarly to Scikit-Learn's OneClassSVM class - a regression-type model, which computes two results:

  1. decisionFunction (continuous double): distance to the separating plane. Inliers have positive values, outliers have negative values. The greater the value, the more typical or atypical, respectively, the data record is.
  2. outlier (categorical boolean): outlier = (decisionFunction < 0 ? true : false).

Example R code:

library("e1071")
library("r2pmml")

data(iris)

iris.simple = svm(~ . - Species, data = iris, type = "one-classification")
r2pmml(iris.simple, "iris-simple.pmml", dataset = iris)

iris.complex = svm(~ . - Species + I(Sepal.Length / Sepal.Width) + I(Petal.Length / Petal.Width), data = iris, type = "one-classification")
iris.complex$xlevels = list(names = c())
r2pmml(iris.complex, "iris-complex.pmml", dataset = iris)

In addition to one-class SVM, the r2pmml package is now able to convert "ordinary" regression and classification SVM as well.

gilad19 commented 7 years ago

Great, thanks a lot!!!

korczykowska commented 5 years ago

Is it possible to apply r2pmml for another type like c-classification?

vruusmann commented 5 years ago

@korczykowska If you have questions, then please open new issue(s), instead of "reusing" old and already closed issues.

Here is a list of supported e1071::svm parameterizations: https://github.com/jpmml/jpmml-r/blob/master/src/test/R/svm.R

By scrolling through this code, I can't tell right away if C-classification is already supported or not (perhaps it's the default for all classifiers). Again, it's your responsibility to provide with a reproducible test case that it's not supported at the moment (sorry, I'm busy working on many things, and I don't have time to verify each claim made against the project).