Heuristic ensembles of filters for accurate and reliable feature selection

Aldehim, Ghadah (2015) Heuristic ensembles of filters for accurate and reliable feature selection. Doctoral thesis, University of East Anglia.

Preview

PDF Download (4MB) | Preview

Abstract

Feature selection has become increasingly important in data mining in recent years. However, the accuracy and stability of feature selection methods vary considerably when used individually, and yet no rule exists to indicate which one should be used for a particular dataset. Thus, an ensemble method that combines the outputs of several individual feature selection methods appears to be a promising approach to address the issue and hence is investigated in this research.
This research aims to develop an effective ensemble that can improve the accuracy and stability of the feature selection. We proposed a novel heuristic ensemble of filters (HEF). It combines two types of filters: subset filters and ranking filters with a heuristic consensus algorithm in order to utilise the strength of each type. The ensemble is tested on ten benchmark datasets and its performance is evaluated by two stability measures and three classifiers. The experimental results demonstrate that HEF improves the stability and accuracy of the selected features and in most cases outperforms the other ensemble algorithms, individual filters and the full feature set.
The research on the HEF algorithm is extended in several dimensions; including more filter members, three novel schemes of mean rank aggregation with partial lists, and three novel schemes for a weighted heuristic ensemble of filters. However, the experimental results demonstrate that adding weight to filters in HEF does not achieve the expected improvement in accuracy, but increases time and space complexity, and clearly decreases stability. Therefore, the core ensemble algorithm (HEF) is demonstrated to be not just simpler but also more reliable and consistent than the later more complicated and weighted ensembles.
In addition, we investigated how to use data in feature selection, using ALL or PART of it. Systematic experiments with thirty five synthetic and benchmark real-world datasets were carried out.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Jackie Webb
Date Deposited:	07 Jun 2016 09:12
Last Modified:	07 Jun 2016 09:12
URI:	https://ueaeprints.uea.ac.uk/id/eprint/59245
DOI:

Downloads

Actions (login required)

View Item