Dear Kasper Christensen,

You may use support vector machine as features selection technqiue to select best set of features. The most simplest way to rank the features using square of the weight assigned by the SVM. Further, you will select log2n of features out of n features. 

This paper is the main one regarding SVM feature selection "Gene Selection for Cancer Classification using Support Vector Machines" http://clopinet.com/isabelle/Papers/geneselect.pdf. I suggest you to start from there. You can apply the method easily using WEKA software.

You may follow these below links,

http://www.jmlr.org/proceedings/papers/v3/chang08a/chang08a.pdf

http://papers.nips.cc/paper/1850-feature-selection-for-svms.pdf

https://www.csie.ntu.edu.tw/~cjlin/papers/features.pdf

Have a look at our paper if you'd like (http://content.iospress.com/articles/intelligent-data-analysis/ida00539). We use the techniques from Recursive Feature Elimination to identify heuristically the most important features in the feature space of a polynomial SVM. The Guyon paper above shows how to identify the most important variables; we go one step further and identify the most important (highest absolute weigth) features. For a degree 3 for example, these features are monomials of degree 3 over the orinal variables. Hope it helps.

Dear Kasper,

Kindly read the links:

https://www.microsoft.com/en-us/research/publication/feature-selection-using-linear-support-vector-machines/

https://www.kaggle.com/c/overfitting/forums/t/487/feature-selection-using-svm

http://www.sciencedirect.com/science/article/pii/S1568494615004093

Dear Kasper Christensen,

SVM is a classification approach based on finding the optimal hyperplan. There is many way to use SVM for feature selection problem:

1. Use SVM to compute the objective function (Classification accuracy rate) and attempt to select the subset of feature that optimizing the classification accuracy rate.

2. Use SVM as filter approach like SVM-RFE, which use the weight vector to select the relevant features.

3. Use the Vapnik–Chervonenkis Bound to compute the important of each feature.