Pattern Classification Program

This release supports model selection for the linear SVM kernel and an option to build SVD transforms using training and test datasets (as opposed to just training data). P-errors are now reported in SVM model selection. The build process was simplified.

This release creates the prediction file pcp.rcl for MLP prediction, implements MLP model selection, implements k-NN model selection, has additional information in the class prediction file pcp.rcl (correct classification flag, TP, FN, FP, and TN flags for two-class cases), removes a major memory handling defect in the forward selection algorithm that lead to poor (computational) performance, enforces the feasible region for nu in NU-SVM, and changes the default number of cross-validation experiments from 10 to 1.

LIBSVM was upgraded to version 2.71. Pearson
correlation was added as a feature selection
criterion. Clustering was removed. The distance
selection menu was removed. Individual class costs
are supported for C-SVM learning. FORTRAN code was
eliminated. A Model Selection menu for the Support
Vector Machine algorithm was added. Forward
selection and backward elimination feature subset
selection algorithms were added. Inter-intra
distance, 1-NN error rate, and Bayes error rate
were added as criteria for feature selection. GNU
autoconf is now used to build PCP.

Ported to Windows (under the Cygwin environment). Minor bugfixes and updated documentation.

The maximum number of attributes for clustering has been increased to 1000. The training data set is used (as opposed to test data set) in clustering. A verbose mode for the clustering summary has been implemented. A formatting bug in saving the results of clustering has been fixed. 'named rows (vectors)' and 'named columns (attributes)' input data file formats are supported. The Golub (ALL/AML) data set is now provided in the more useful named rows/named columns format.