Principal Direction Linear Oracle for Gene Expression Ensemble Classification

A principal direction linear oracle (PDLO) ensemble classifier for DNA microarray gene expression data is proposed. The common fusion-selection ensemble based on weighted trust for a specifier classifier was replaced with pairs of subclassifiers of the same type using PDLO to perform a linear hyperplane split of training and testing samples. The hyperplane split forming the oracle was based on rotations of principal components extracted from sets of filtered features in order to maximize the separation of samples between the pair of miniclassifiers. Eleven classifiers were evaluated for performance with and without PDLO implementation, which included k nearest neighbor (kNN), naı̈ve Bayes classifier (NBC), linear discriminant analysis (LDA), learning vector quantization (LVQ1), polytomous logistic regression (PLOG), artificial neural networks (ANN), constricted particle swarm optimization (CPSO), kernel regression (KREG), radial basis function networks (RBFN), gradient descent support vector machines (SVMGD), and least squares support vector machines (SVMLS). PLOG resulted in the best performance when used as a base classifier for PDLO. The greatest performance for PLOG implemented with PDLO occurred for tenfold CV and 100 rotations of PC scores with fixed angles for hyperplane splits. Random rotation angles for hyperplane splits resulted in reduced performance when compared to rotations with fixed angles.