Experimental Study of Different FSAs in Classifying Protein Function

Abstract

This paper addresses one of the challenges of machine learning in improving performance through feature selection algorithms (FSAs). Application of FSAs in the bioinformatics domain has become a necessity due to enormous growth of public sequence databases. This paper provides an experimental framework on the use of Rough Set Theory (RST) as FSAs in finding minimal feature subsets for classifying protein function. In experimenting RST, three different recent models are explored; Correlation Feature Selection (CFS), FCBF (Fast Correlation-Based Filter) and Artificial Immune System (AIS). The experimental study for these FSAs are based on four criteria: the accuracy (AC), the area under ROC graph (ROC), the length of the reducts (ARL), and the time taken (TT). Classification was performed on the reduced feature set using the Support Vector Machine algorithm. The results demonstrate that CFS and FCBF performs better if the main objectives are to measure the accuracy and ROC, however in terms of duration and rule length, RST is a better choice.