Discipline: Technology and Engineering
Subcategory: Electrical Engineering
Keenan Leatham - University of the District of Columbia
Co-Author(s): Saul Henderson, Nian Zhang and Lara Thompson
Recent research has showed great success of a combination of PSO and least squares support vector machines (LS-SVM) to solve feature selection problem, where the LS-SVM was used as a classifier. They use fewer features and takes shorter computational time than other feature selection algorithms. Because the PSO-EA algorithm has shown evident advantage including global optimization and much higher classification accuracy over the PSO algorithm according to our prior work, it is anticipated that the proposed PSO-EA algorithm will obtain better performance than current PSO algorithm on feature selection problem. So far, no work has been found to use the PSO-EA algorithm with LS-SVM classifier for feature selection problem. Therefore, it is imperative to develop a simultaneous parameter optimization and feature selection algorithm using the hybrid particle swarm optimization (PSO) and evolutionary algorithm (EA) with an effective least squares support vector machine (LS-SVM) classifier to evaluate the performance of the feature subset. First, a novel hybrid PSO-EA model is built with the (Gaussian) RBF function be the kernel function. Next, the fitness function will be determined. The fitness function represents how well the ith particle’s position in the multidimensional space is relatively to the desired goal. The fitness function will be carefully designed to meet the classification accuracy and minimum feature number objectives. It could be the mean square root error of the training samples. Then a hybrid particle swarm optimization and evolutional (PSO-EA – LS-SVM) algorithm will be developed. The algorithm will not only minimize the number of features, but also help minimize the classification error rate for the LS-SVM classifier. The result will be a LS-SVM model at the optimal parameters, alpha and sigma, feature subset, and classification accuracy. Once the alpha and sigma parameters are tuned, a training algorithm will be designed to train the support values and the bias term for function approximation. 80% of the data will be used to train the network, and the remaining 20% data points will be used as the testing data. After the network has been well trained, we use the LS-SVM classifier to evaluate the performance of the new feature subsets, which have never been seen by the network. Finally, a comprehensive comparison among the proposed PSO-EA – LSSAM model, CSA – LS-SVM model, PSO-LS-SVM, and EA-LS-SVM will be performed in terms of classification accuracy, sensitivity (i.e. true positive rate), and specificity (i.e. true negative rate). The success rate of the proposed PSO-EA – LSSAM model is over 97%, which outperform other competing approaches. The findings demonstrate the effectiveness and high efficiency of the developed method.
Funder Acknowledgement(s): National Science Foundation (NSF/HBCU-UP/ HRD #1505509, HRD #1533479, and NSF/DUE #1654474)
Faculty Advisor: Nian Zhang, nzhang@udc.edu
Role: I analyzed the behavior of simultaneous parameter optimization and feature selection algorithm using the hybrid particle swarm optimization.