Basic classifiers optimised for LOOCV
Matlab Fortran M
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
+looc
+looc_sorted
.gitignore
LICENSE
README.md
make_mex.m

README.md

loo_classifiers

This is a Matlab package to efficiently implement leave-one-out cross-validation for a number of basic supervised learning algorithms.

I was surprised I couldn't find anything online to do this, so perhaps this will be useful to other people. There are two Matlab packages implementing different classifiers.

The first, looc_sorted assumes balanced data (same number of trials for every class) and uniform priors over the classes. The input data should be already sorted by class. All the algorithms in this package take as the first argument a 3d data array with dimensions (Nftr, Ncls, Ntrl), and returns two arguments - the confusion matrix (Ncls, Ncls) and the information in the confusion matrix. Some classifiers have extra parameters which should be passed as additional input arguments.

The second package is looc which drops the requirement for balanced inputs and has an input format of X with dimension (Ntrl, Nftr) being the feature values and Y with dimension (Ntrl,1) being the class labels (integers starting from 1). Uniform priors over classes are still assumed (even if there are different numbers of trials per class in the input).

Where possible, results were tested against Matlab classify from the Statistics Toolbox.

Algorithms

The currently implemented algorithms are:

  • diag_linear : MVN clusters, pooled diagonal covariance matrix
  • linear : MVN clusters, pooled covariance matrix
  • diag_quadratic : MVN clusters, diagonal covariance matrices
  • quadratic : MVN clusters
  • nearest_mean : MVN clusters, pooled diagonal covariance, equal variances (template matching)
  • poisson_bayes : Independent poisson features for each cluster. Count data only
  • multinom_bayes : Independent multinomial features for each cluster. Count data only
  • knn : k-Nearest neighbour classifier.

Contact

Bugs / comments / ideas / criticism to robince at gmail dt com

This is unpublished research code so comes with no guarentees. Please feel free to try it out, but if it is useful to you and forms a key part of an analysis please consider contacting me. Perhaps I could help in some way with feedback or modifications or optimisations specific to your case. I don't have any citation yet to provide for this but will update here if/when I do.

License

This project is licensed under version 3 of the GNU General Public License. For the exact terms please see the LICENSE file.

ToDo

  • add more classifiers to looc
  • provide versions / option to drop uniform priors and take priors from data