An improved functional link neural network for data classification
The goal of classification is to assign the pre-specified group or class to an instance based on the observed features related to that instance. The implementation of several classification models is challenging as some only work well when the underlying assumptions are satisfied. In order to genera...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/266/1/24p%20UMER%20IQBAL.pdf http://eprints.uthm.edu.my/266/2/UMER%20IQBAL%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/266/3/UMER%20IQBAL%20WATERMARK.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The goal of classification is to assign the pre-specified group or class to an instance based on the observed features related to that instance. The implementation of several classification models is challenging as some only work well when the underlying assumptions are satisfied. In order to generate the complex mapping between input and output space to build the arbitrary complex non-linear decision boundaries, neural networks has become prominent tool with wide range of applications. The recent techniques such as Multilayer Perceptron (MLP), standard Functional Link Neural Network (FLNN) and Chebyshev Functional Link Neural Network (CFLNN) outperformed their existing regression, multiple regression, quadratic regression, stepwise polynomials, K-nearest neighbor (K-NN), Naïve Bayesian classifier and logistic regression. This research work explores the insufficiencies of well- known CFLNN model where CFLNN utilizes functional expansion with large number of degree and coefficient value for inputs enhancement which increase computational complexity of the network. Accordingly, two alternative models namely; Genocchi Functional Link Neural Network (GFLNN) and Chebyshev Wavelets Functional Link Neural Network (CWFLNN) are proposed. The novelty of these approaches is that, GFLNN presents the functional expansions with less degree and small coefficient values to make less computational inputs for training to overcome the drawbacks of CFLNN. Whereas, CWFLNN is capable to generate more number of small coefficient value based basis functions with same degree of polynomials as compared to other polynomials and it has orthonormality condition therefore it has more accurate constant of functional expansion and can approximate the functions within the interval. These properties of CWFLNN are used to overcome the deficiencies of GFLNN. The significance of proposed models is verified by using statistical tests such as Freidman test based on accuracy ranking and pairwise comparison test. Moreover, MLP, standard FLNN and CFLNN are used for comparison. For experiments, benched marked data sets from UCI repository, SVMLIB data set and KEEL data sets are utilized. The CWFLNN reveals significant improvement (due to its generating more numbers of basis function property) in terms of classification accuracy and reduces the computational work. |
|---|