Information measure toolbox for classifier evaluation on open source software Scilab

This paper describes an information measure toolbox for classifier evaluations based on an open-source software platform Scilab. Twenty four normalized information measures are given in the toolbox, which are derived from three types of information definitions, namely, mutual information, information divergence, and cross entropy. Different from the conventional performance measures which apply the heuristic or empirical formulas, the information measures are more theoretically sound with a higher degree of applicability, say, classifications including a reject option. The specific attention is paid to the singularity aspect in the implementation of the toolbox. With the toolbox, users are able to test the included numerical examples on binary and three-class classifications easily. The toolbox developed in this work provides users a useful tool of assessing classifiers from an information theoretic basis. The complete source code of the toolbox is available at website ¿OpenPR¿: http://www.openpr.org.cn/ with a file name of ¿confmatrix2ni.zip¿.