Comparison of Heuristics for Inhibitory Rule Optimization

Abstract
Knowledge representation and extraction are very important tasks in data mining. In this work, we proposed a variety of rule-based greedy algorithms that able to obtain knowledge contained in a given dataset as a series of inhibitory rules containing an expression “attribute ≠ value” on the right-hand side. The main goal of this paper is to determine based on rule characteristics, rule length and coverage, whether the proposed rule heuristics are statistically significantly different or not; if so, we aim to identify the best performing rule heuristics for minimization of rule length and maximization of rule coverage.

Friedman test with Nemenyi post-hoc are used to compare the greedy algorithms statistically against each other for length and coverage. The experiments are carried out on real datasets from UCI Machine Learning Repository. For leading heuristics, the constructed rules are compared with optimal ones obtained based on dynamic programming approach. The results seem to be promising for the best heuristics: the average relative difference between length (coverage) of constructed and optimal rules is at most 2.27% (7%, respectively). Furthermore, the quality of classifiers based on sets of inhibitory rules constructed by the considered heuristics are compared against each other, and the results show that the three best heuristics from the point of view classification accuracy coincides with the three well-performed heuristics from the point of view of rule length minimization.

Citation
Comparison of Heuristics for Inhibitory Rule Optimization 2014, 35:378 Procedia Computer Science

Publisher
Elsevier BV

Journal
Procedia Computer Science

Conference/Event Name
International Conference on Knowledge-Based and Intelligent Information and Engineering Systems, KES 2014

DOI
10.1016/j.procs.2014.08.118

Additional Links
http://linkinghub.elsevier.com/retrieve/pii/S1877050914010837

Permanent link to this record