Error-Rate Evaluation of Classification Data Mining Algorithms in Multidisciplinary Educational Data

Abstract

Olamiti A. O. Osofisan A. O.
aolamiti@yahoo.com nikeosofisan@gmail.com
Corresponding Author
Department of Computer Science,
University of Ibadan, Ibadan, Nigeria
Abstract
Comparisons tests on Knowledge Discovery in Data (KDD) methods, techniques and tools are carried out to improve on them
and also to come up with those that are believed to be “best” for specific domains. Comparing the absolute difference of the
error-rates of the algorithms is not enough because the difference should also be tested statistically. Various statistical tests are
thus used to determine models/classifiers performances. This study evaluated the performance of the two mostly adopted
educational data mining algorithms namely Classification And Regression Trees (CART) and C4.5 with Educational Data (ED)
which has the specific characteristic of normal class distribution. The CART and C4.5 were used independently to build ten
models for ten ratios. The CART and C4.5 error-rate averages were calculated and their classification performances were
compared using two-tailed t-test at ?0.05. The difference in the error-rates of CART and C4.5 is shown to be statistically
significant.