Normalize data for libsvm

The section "3.2.5. Tips on Practical Use" on this page http://scikit-learn.org/stable/modules/svm.html recommends to scale the data before using. This document doesn't tell anything about internal scaling. I don't think the scikit-learn library will scale normalize the data internally. Here is the document which the above page refers to http://scikit-learn.org/stable/modules/preprocessing.html#preprocessing

You have to do scaling or normalization separately. The preprocessing module has various classes and function for this:

• MinMaxScaler scales individual features so that the min and max in the training become -1 and 1, respectively
• StandardScaler removes the mean, then scales to unit variance (per feature)
• Normalizer normalizes entirely samples to unit length according to the L1 or L2 norm

Normalization is appropriate when handling things like term frequencies. For simple Gaussian or nearly-Gaussian features, use scaling.

Pipeline objects make these a lot easier to use:

svm = Pipeline([('scale', MinMaxScaler()),
                ('svm', SVC(kernel="rbf"))])