Feature Analysis and Classification of Protein Secondary Structure Data

Abstract

In this paper, we investigate feature analysis for the prediction of the secondary structure of protein sequences using support vector machines (SVMs) and k-nearest neighbor algorithm (kNN). We apply feature selection and scaling techniques to obtain a number of distinct feature subsets with different features and each scaled differently. The feature selection and the scaling are performed using the mutual information (MI). We formulate the feature selection and scaling as combinatorial optimization problem and obtain solutions using a Hopfieldstyle algorithm. Our experimental results show that the feature subset selection improves the performance for both SVM and kNN while the feature scaling is consistently beneficial for kNN.KeywordsSupport Vector MachineFeature SelectionMutual InformationEnergy FunctionFeature SubsetThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.