******** Start here ****************************************************
**************************************************************************
Try to run first example:
>> load('benchmarks.mat')
>> launch('titanic', titanic, 'SVMl1', 'linear', 1)
Try to run second example:
>> load('benchmarks.mat')
>> launch('thyroid', titanic, 'RMM', 'linear', 1)
%% Note that to use RMM, you need to install MOSEK toolbox.
**************************************************************************
**************************************************************************
1. To use the RMM code [Shivaswamy&Jebara 2010], you need to install MOSEK optimization toolbox.
2. Data is taken from [Shivaswamy&Jebara 2010] and [Gunnar Ratsch et al 2001], more can be downloaded at
http://theoval.cmp.uea.ac.uk/~gcc/matlab/default.html#benchmarks
3. Data used in the second set of experiments: breast_cancer, heart, thyroid, titanic.
For random split i , train the model with, for example, heart.x(train(i,:), :) and heart.t((train(i,:), :)) and test with, for example, heart.x(test(i,:), :) and heart.t((test(i,:), :))
4. For the first set of experiment, 10 fold cross validation was used, with inner 10 fold cross validation to choose optimal hyperparams.
For the second set of experiment, 100 random splits (default, the same as in [Shivaswamy&Jebara 2010] and [Ratsch2001]) was used, with inner 5 fold cross validation to choose optimal hyperparams.
5. For SVMRFE, use function CV_SVMRFE_CVNumFeat().
6. For other algorithms, use function launch().
7. For loading results (to take average of 100 splits), use function load_split_results().
8. For statistical test, using t-test, use function statisticalTest().
9. For statistical test, using McNemar test, use function mcnemar(). Note that McNemar fits for k-fold cross validation, not 100 random splits.
