Open Access Repository

Novel feature selection algorithms for improving neural network performance

Zhao, Z 2017 , 'Novel feature selection algorithms for improving neural network performance', PhD thesis, University of Tasmania.

Full text not available from this repository.

Abstract

Data mining and machine learning have become enormously pivotal in this Big Data
time, as people are tremendously eager to predict from what they have known, and
foresee the unknown. During the process of machine learning, selecting features that are
genuinely useful and helpful to the prediction tasks remains a central issue, because a
smaller but more relevant dataset can improve the performance of machine learning.
In this thesis, the whole procedure of feature selection is investigated, beginning with a
new data preparing method, Average Random Choosing Method (ARCM), which can
solve Class Imbalance Problem. Experimental results show that ARCM can
significantly improve the prediction accuracy of Artificial Neural Network (ANN) for
imbalanced datasets, in comparison with other researches’ results.
After that, the new filter named Consistency Concentration Based Feature Selection
(CCBFS) is introduced. It is based on the conception of consistency based feature
selection (CBFS). CCBFS can evaluate each individual feature and works with both
nominal and continuous features, which used to be a shortage of consistency based
algorithms.
A GA based wrapper is proposed after CCBFS. This Accumulate Elitism Genetic
Algorithm (AEGA) searching approach for feature selection inherits the advantages of
GA based methods: quick parallel searching and large searching space. AEGA also
covers the limitation of GA such as slow converging speed. Especially for ANN, AEGA
has good resistance on the unstable performance of MLP and shortens the training times.
In the end, the hybrid structure Multi Combined Filter-Wrapper Feature Selection
(MCFWFS) is introduced. This novel algorithm uses CCBFS to make a pre-selection
and accelerate converging process for AEGA. The multi combined structure ensures fast
computing time and large improvement of predictor’s performances at the same time. It
is designed for a wide range of real world problems, even with extremely large datasets.
The main contributions of this research include a novel data preparing method (ARCM),
a novel filter for feature ranking (CCBFS), a novel GA based wrapper (AEGA) and a
novel structure of hybrid feature selection (MCFWFS). Experiments have demonstrated
that these new algorithms have outstanding performances on many kinds of datasets and
real world problems in the Machine Learning categories of classification and multi class
recognition. This research provides new solutions for improving the machine learning
performances and make up shortages of current feature selection methods at some
extent.

Item Type: Thesis - PhD
Authors/Creators:Zhao, Z
Keywords: Feature Selection, Artificial Neural Network, Machine Learning, Data Mining,
Copyright Information:

Copyright 2016 the Author

Item Statistics: View statistics for this item

Actions (login required)

Item Control Page Item Control Page
TOP