ajaysapkota / tsvm

experiments testing transductive svm for my blog posts

tsvm

The Goal of this project is to design and run data science experiments to test various transductive and semi-supervised learning algorithms

The TSVM theory is described on my blog http://charlesmartin14.wordpress.com/2014/07/06/machine-learning-with-missing-labels-transductive-svms/

The first objective is to test svmlin http://vikas.sindhwani.org/svmlin.html

against liblinear http://www.csie.ntu.edu.tw/~cjlin/liblinear/

using the binary datasets provided for libsvm http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html

to determine how to tune svmlin well and what kind of data sets it performs well on

Later, we would like to look at Semi-Supervised learning algos such as

http://www.dii.unisi.it/~melacci/lapsvmp/

and the python scikit learn label propagation algo

http://scikit-learn.org/stable/modules/label_propagation.html

For Newbies: If you don't know anything about machine learning, you should first learn how to run liblinear on the libsvm data sets

We need someone to create a liblinear tutorial For now, you can see http://jamescpoole.com/2012/10/30/libsvm-tutorial-part-1-overview/

libsvm is almost identical to liblinear

To get started 0. requirements: ruby 2.x and gnu parallel ruby can be installed using rvm gnu parallel should be in the path

1. download and install liblinear and svmlin

2. download the a1a trainig and test data sets

http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/a1a

http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/a1a.t

3. edit svmlin, set the variables

SVMLIN_DIR = "~/packages/svmlin-v1.0"

LIBLINEAR_DIR = "~/packages/liblinear-1.94"

4. run svmlin.rb a1a

5. repeat for the a2a, a3a, ... data sets and the w2a, w3a, ... data sets