Publication

Abstract : In todays “data-centric” world, the prevalence of vast and immeasurable amount of data pertaining to various fields of study has led to the need for properly analyzing and apprehending this information to yield knowledge that becomes useful in decision making. Among the many procedures for handling this multitude of data, “classification” is the one that aids in making decisions based on categorization of data and “feature selection” is the process of picking out attributes relevant to the study. Keeping classification as the central idea of our study, we aim at presenting a comparative analysis of prediction accuracies obtained by two chosen classification algorithms, namely, SVM and RBFN. We proceed to introduce feature selection using both filter and wrapper methods along with SVM and RBFN to showcase a detailed analytical report on variations in performance when using classification algorithms alone, and with application of feature selection. The four approaches used for feature selection in our study are; Information Gain, Correlation, Particle Swarm Optimization (PSO) and Greedy method. Performance of the algorithms under study is evaluated based on time, accuracy of prediction and area under ROC curve. Although time and accuracy are effective parameters for comparison, we propose to consider ROC area as the criterion for performance evaluation. An optimal solution will have the area under ROC curve value approaching 1.