International Journal of Operations Research and Information Systems, 6(4), 1-18, October-December 2015 1
Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
Keywords: Classifcation, Component Design, Data Mining, Decision Tree, Securities Account Holders
ABSTRACT
Popular decision tree (DT) algorithms such as ID3, C4.5, CART, CHAID and QUEST may have different
results using same data set. They consist of components which have similar functionalities. These components
implemented on different ways and they have different performance. The best way to get an optimal DT for a
data set is one that use component-based design, which enables user to intelligently select in advance imple-
mented components well suited to specifc data set. In this article the authors proposed component-based design
of the optimal DT for classifcation of securities account holders. Research results showed that the optimal
algorithm is not one of the original DT algorithms. This fact confrms that the component design provided
algorithms with better performance than the original ones. Also, the authors found how the specifcities of
the data infuence the DT components performance. Obtained results of classifcation can be useful to the
future investors in the Montenegrin capital market.
Component-Based
Decision Trees:
Empirical Testing on Data Sets of Account
Holders in the Montenegrin Capital Market
Ljiljana Kašćelan, Faculty of Economics, University of Montenegro, Podgorica, Montenegro
Vladimir Kašćelan, Faculty of Economics, University of Montenegro, Podgorica, Montenegro
1. INTRODUCTION
Popular DT algorithms are implemented with “black-box”-approach, which implies that the
logic is hidden from users and there is no possibility of ad hoc changes. These algorithms are
changed and improved incrementally, and that can take a long time.
Component-based design implies “white-box”-construction of algorithms with help of stan-
dardized reusable components (RCs) obtained from original DT algorithms (Delibasic, Jovanovic,
Vukicevic, Suknovic & Obradovic, 2011; Suknovic, Delibasic, Jovanovic, Vukicevic, Becejski-
Vujaklija & Obradovic, 2012). It enables combination of advantages of various algorithms and
their comparison, as well as the testing of influence of certain components on performance of
the algorithms. Combining these components we can improve performance and get optimal DT
algorithm for a specific data set.
DOI: 10.4018/IJORIS.2015100101