Preceding rule induction with instance reduction methods

Othman, Osama; Bryant, CH

Abstract

A new prepruning technique for rule induction is presented which applies instance reduction before rule induction. An empirical evaluation records the predictive accuracy and size of rule-sets generated from 24 datasets from the UCI Machine Learning Repository. Three instance reduction algorithms (Edited Nearest Neighbour, AllKnn and DROP5) are compared. Each one is used to reduce the size of the training set, prior to inducing a set of rules using Clark and Boswell's modification of CN2. A hybrid instance reduction algorithm (comprised of AllKnn and DROP5) is also tested. For most of the datasets, pruning the training set using ENN, AllKnn or the hybrid significantly reduces the number of rules generated by CN2, without adversely affecting the predictive performance. The hybrid achieves the highest average predictive accuracy.

• Presentation Conference Type: Conference Paper (published)
• Publication Date: Jul 1, 2013
• Deposit Date: Aug 8, 2013
• Publicly Available Date: Apr 5, 2016
• Publisher: Springer
• Pages: 209-218
• Series Title: Lecture Notes in Computer Science
• Series Number: 7988
• Book Title: Proceedings of the 9th International Conference on Machine Learning and Data Mining in Pattern Recognition.
• ISBN: 9783642397110
• DOI: https://doi.org/10.1007/978-3-642-39712-7_16
• Keywords: Rule Induction, Overfitting, Noise Filtering, Instance Reduction
• Related Public URLs: http://www.salford.ac.uk/computing-science-engineering/cse-academics/chris-bryant; http://www.springer.com/; http://www.springer.com/computer/ai/book/978-3-642-39711-0; http://www.mldm.de/
• Additional Information: Additional Information : MLDM 2013 was held between 19-25 July 2013 in New York, USA.