Can one please refer me to a formalism and optimization algorithms of SVM which outputs not only 0/1 but a vector of 0/1's, i.e. I'd like to train multiple classifiers at once for the same input.
Thanks!
1
-
$\begingroup$ You kinda answered your own question: train multiple classifiers. If you want some kind of structured output, such as a graph, you need different techniques. $\endgroup$– Marc ClaesenCommented Oct 9, 2013 at 8:33
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
Most SVM implementations offer strategies or methods for multi-label classification; libsvm has details here, and libraries like scikit-learn (Python) have wrappers and classifiers for this. You can also train a SVM model to output probability estimates or distance measures for each class and use thresholding to apply multiple labels.
$\begingroup$
$\endgroup$
1
Actually you can train an SVM which can output vectors and more. As was noted by Marc Claesen such outputs are called structured output, and are formulated in terms of structured output SVM. It was introduced by Tsochantaridis et. al. (2004). Here is the link to the project which contains code and examples: structured output project page
-
$\begingroup$ Could somebody elaborate why the answer was down voted? $\endgroup$– GnattuhaCommented Apr 24, 2014 at 0:02