What's the difference between using libSVM in sci-kit learn, or e1070 in R, for training and using support vector machines?

Question

Recently I was contemplating the choice of using either R or Python to train support vector machines.

Aside from the particular strengths and weaknesses intrinsic to both programming languages, I'm wondering if there is any heuristic guidelines for making a decision on which way to go, based on the packages themselves.

I'm thinking in terms of speed of training a model, scalability, availability of different kernels, and other such performance-related aspects.

Given some data sets of different sizes, how could one decide which path to take?

I apologize in advance for such a possibly vague question.

Answer 1

I do not have experiece with e1070, however from googling it it seems that it either uses or is based on LIBSVM (I don't know enough R to determine which from the cran entry). Scilearnkit also uses LIBSVM.

In both cases the model is going to be trained by LIBSVM. Speed, scalability, variety of options available is going to be exactly the same, and in using SVMs with these libraries the main limitations you will face are the limitations of LIBSVM.

I think that giving further advice is going to be difficult unless you clarify a couple of things in your question: what is your objective? Do you already know LIBSVM? Is this a learning project? Who is paying for your time? Do you feel more comfortable in Python or in R?

Answer 2

Sometime back I had the same question. Yes, both e1070 and scikit-learn use LIBSVM. I have experience with e1070 only.

But there are some areas where R is better . I have read in the past that Python does not handle categorical features properly (at least not right out of the box). This could be a big deal for some.

I also prefer R's formula interface. And some of the nice data manipulation packages.

Python is definitely better for general purpose programming and scikit-learn aids in using a single programming language for all tasks.