Improved Bayesian Feature Selection and Classification Methods Using Bootstrap Prior Techniques
No Thumbnail Available
Date
2016
Journal Title
Journal ISSN
Volume Title
Publisher
Faculty of Computer and Applied Computer Science, Tibiscus University of Timisoara, Romania
Abstract
In this paper, the behavior of feature
selection algorithms using the traditional t-test, Bayesian
t-test using MCMC and Bayesian two-sample test using
proposed bootstrap prior technique were determined. In
addition, we considered some frequentist classification
methods like k- Nearest Neighbor (k-NN), Logistic
Discriminant (LD), Linear discriminant analysis (LDA),
Quadratic discriminant analysis (QDA) and Naïve Bayes
when conditional independence assumption is violated.
Two new Bayesian classifiers (B-LDA and B-QDA) were
developed within the frame work of LDA and QDA using
the bootstrap prior technique. The model parameters were
estimated using Bayesian approach via the posterior
distribution that involves normalizing the prior for the
attributes and the likelihood from the sample in a MonteCarlo experiment. The bootstrap prior technique was
incorporated into the Normal-Inverse-Wishart natural
conjugate prior for the parameters of the multivariate
normal distribution where the scale and location
parameters were required. All the classifiers were
implemented on the simulated data at 90:10 training-test
data ratio. The efficiencies of these classifiers were
assessed using the misclassification error rate, sensitivity,
specificity, positive predictive value, negative predictive
value and area under the ROC curve. Results from various
analyses established the supremacy of the proposed Bayes
classifiers (B-LDA and B-QDA) over the existing
frequentists and Naïve Bayes classification methods
considered. All these methods including the proposed one
were implemented on a published binary response
microarray data set to validate the results from the
simulation study
Description
Keywords
k-Nearest Neighbour, Bayesian Linear Discriminant Analysis, Bayesian Quadratic Discriminant Analysis, Naïve Bayes, Bootstrap prior