Release notes

Gist Release 2.3, March 22, 2006

• Add the -algorithm lin option to gist-sigmoid.
• Fixed the diffusion kernel in gist-kernel.
• Added the int-dim, simple-diffusion and modav operations to gist-matrix.

Gist Release 2.2, Sep 20, 2005

• Added the -hyperplane option to 'gist-train-svm' and added the corresponding 'gist-fast-classify' program.
• Added the gist-kernel program.

Gist Release 2.1, May 3 2004

• The main difference between 2.1 and 2.0.9 is the renaming of the executables. This means that 'compute-weights' is now 'gist-train-svm' and 'classify' is 'gist-classify'. The new executable names are documented here.

Gist Release 2.0.9, May 3 2004

• The Gist system is now built from code using the GNU build system. This should make it easier for users on diverse systems use or adapt our code.
• Minor changes to code to allow compilation on Windows (Cygwin) and Macintosh OSX. Binaries for these platforms have been added to the distribution options.
• Added -precision option to control the number of digits after the decimal point in output files.
• Included more informative error messages when trying to do feature selection on data that has too few positive or negative training examples.
• Fixed a minor bug in the -widthfactor option to compute-weights and kernel-pca. Previously, contrary to what the documentation said, the width factor F was applied after squaring the width w in the radial basis kernel. This has been fixed, so that F is multiplied by w prior to squaring.
• Added the -width option to compute-weights and kernel-pca, allowing the user to set the radial basis width directly.
• gist-rfe now retains the sign on the feature weights, so that you can tell whether the feature occurs more frequently in the positive or negative class.

Gist Release 2.0.8, January 6, 2004

• classify: Bugs fixed in matrix dimension tests for "self-train" and "self-test" matrices.
• compute-weights: When -kernelout is selected, the diagonal factor for the 2-norm soft margin is no longer added. This change makes it possible to use compute-weights to generate kernel matrices that can be modified externally, and then re-entered for training using the -matrix option.
• fselect: The -threshold option has been renamed to -fthreshold to agree with compute-weights.

Gist Release 2.0.7, December 16, 2003

• Changed the default value of 'maxiter' (maximum number of iterations during training) to 'no limit', from the previous default of 10000. This prevents early termination of jobs that take many iterations. Users who want iteration or time limits should set the -maxiter or -maxtime options for compute-weights.
• Fixed minor bug in the array shuffling routine that caused arrays to be non-uniformly shuffled.
• Default precision and width in rdb-matrix increased from 5/3 to 8/5. This reduces problems caused by loss of precision during manipulation of files.
• Added a Perl script, gist-rfe to do SVM recursive feature elimination.

Gist Release 2.0.6 was not released.

Gist Release 2.0.5, April 30, 2003

• compute-weights: Fixed bug in 2-norm soft margin implementation that caused incorrect use of the diagonal factors in some cases. In particular, training errors could be mis-reported in some cases, and hold-one-out cross-validation could yield incorrect results in some situations. This only affects applications where the 2-norm soft margin was used (that is, with a 'diagonal factor').
• score-svm-results: Uninitialized variable bug fixed.

Gist Release 2.0.4, April 3, 2003

• classify: Fixed bug that caused failure when using a non-square test kernel matrix.
• classify: Using a training kernel matrix for 'classify' when not needed causes a warning to be issued.
• Got rid of dependency on system utilities 'hostname' and 'date'; UTIL_PATH is no longer needed.

Gist Release 2.0.3, February 20, 2003

• Modification of how system utilities are found; there is a UTIL_PATH define that can now be set in the makefile if 'hostname' and 'date' are not in the user's path.
• Documentation improvements.

Gist Release 2.0.2, October 11, 2002

• Modify fit-sigmoid so that it takes three inputs: a training set of predictions for fitting the sigmoid, the corresponding labels, and a test set of predictions that are to be converted to probabilities. Note that, for the curve fit to be valid, the sigmoid training set should be distinct from the set used to train the SVM.
• Fixed bug in normal approximation to Mann-Whitney U test p-value calculation.

Gist Release 2.0.1, August 14, 2002

• Modify compute-weights so that the default parameters are reasonable: use a 2-norm soft margin with a diagonal factor of 0.1; turn normalization on; use a constant factor of 10.

Gist Release 2.0, July 24, 2002

• Rename the entire package.
• Add kernel-pca, project, fit-sigmoid, fselect and score-svm-results programs.
• Add many feature selection options to compute-weights.
• Add many options to rdb-matrix.
• Improve I/O error checking.
• Generalize label file format to allow multiple classes in one file.

svm Release 1.2, August 21, 2001

• Add -rdb option to compute-weights and classify, and make the default behavior to not expect or produce RDB format lines.
• Fix bug in shuffling routine used in selecting order of evaluation of examples during training and hold-one-out testing. This change will not affect the output of the program, but may affect the speed of convergence.
• Add -maxiter option to compute-weights.
• Add -holdout option to compute-weights.
• Scale precision of weights output depending upon convergence threshold.
• Add -initial option to compute-weights.

svm Release 1.1, July, 2000.

• Add -adddiag option to compute-weights.
• Add -zeromean and -varone options to compute-weights.
• Add a secondary optimization of only the support vectors.
• Make the convergence threshold depend upon the number of examples.

svm Release 1.0, September 23, 1999.

• Fix bug in weight update routine.
• Include median kernel value in diagonal factor heuristic.
• Add the constant factor back in after diagonalizing.
• Fix bug in the way polynomialization was done.

svm Release 0.95, August 30, 1999.

• Change the normalization routines to operate on the kernel, rather than the data. This makes the normalization appropriate for any kernel.
• Generalize the radial kernel to work as a transformation of any kernel function.
• Remove the covariance kernel and the sigmoid kernel.
• Switch the way polynomialization works. It is now A(K + B))^C, rather than ((AK) + B)^C.

svm Release 0.9, August 16, 1999.