A Bootstrap-based Linear Classifier Fusion System for Protein Subcellular Location Prediction

Abstract

The subcellular location plays a pivotal role in the functionality of proteins. In this paper we develop a multi-stage linear classifier fusion system based on Efron's bootstrap sampling for predicting subcellular locations of yeast proteins. Three different types of classifiers, i.e. the Naive Bayes (NB) classifier, radial basis function (RBF) network, and multilayer perceptron (MLP), are utilized to construct the component modules in the fusion system. Ten bootstrapped instance sets are generated for training each type of component classifiers respectively. The linear fusion models, updated by the least-mean-square (LMS) algorithm, are used to integrate the local decisions of the component classifiers and derive the final predictions. The empirical results show that the RBF classifiers can reach at slightly higher accuracy and better precision versus the NB or MLP ones. The linear fusion system consistently improves the overall prediction accuracy, in particular 6.65%, 1.77%, and 3.21%, superior to the NB, RBF, and MLP component classifiers, respectively.