Statistical validation of multiple classifiers over multiple datasets in the field of pattern recognition

Classification is a machine learning technique which is used to categorise the different input patterns into different classes. To select the best classifier for a given dataset is one of the critical issues in pattern classification. Using cross-validation approach, it is possible to apply candidate algorithms on a given dataset and best classifier is selected by considering various evaluation measures of classification. But computational cost associated with this approach is trivial in nature. There are a lot of tests described in the literature but they are specifically used on different and non-related datasets. But, in pattern recognition domain the datasets are taken from the same testing samples and hence are related in nature. In this paper, we propose the use of two non-parametric tests, namely Mann-Whitney U test for comparison of two classifiers and Kruskal-Wallis H test with the corresponding post-hoc tests for comparison of multiple classifiers on multiple datasets. The tests have been fruitfully applied on the datasets of handwritten digit recognition problem taken from UCI Machine Learning Database Repository (http://www.ics.uci.edu/~mlearn). The results show that both the tests provide a better solution than previously proposed tests.