An example of using feature selection with a classifier

For CoSMoMVPA's copyright information and license terms, #
see the COPYING file distributed with CoSMoMVPA. #

% This example plots classification accuracy as a function of the number of
% features selected using an ANOVA on training set data
config = cosmo_config();
data_path = fullfile(config.tutorial_data_path, 'ak6', 's01');

data_fn = fullfile(data_path, 'glm_T_stats_perrun.nii');
mask_fn = fullfile(data_path, 'vt_mask.nii');
ds = cosmo_fmri_dataset(data_fn, 'mask', mask_fn, ...
                        'targets', repmat(1:6, 1, 10), ...
                        'chunks', floor(((1:60) - 1) / 6) + 1);

% remove constant features
ds = cosmo_remove_useless_data(ds);

opt = struct();
opt.child_classifier = @cosmo_classify_naive_bayes;
opt.feature_selector = @cosmo_anova_feature_selector;

partitions = cosmo_nfold_partitioner(ds);

ratios_to_keep = .05:.05:.95;
nratios = numel(ratios_to_keep);

accs = zeros(nratios, 1);

for k = 1:nratios
    opt.feature_selection_ratio_to_keep = ratios_to_keep(k);

    [pred, acc] = cosmo_crossvalidate(ds, ...
                                      @cosmo_meta_feature_selection_classifier, ...
                                      partitions, opt);
    accs(k) = acc;
end

plot(ratios_to_keep, accs);
xlabel('ratio of selected feaures');
ylabel('classification accuracy');