BCBCSF: Bias-corrected Bayesian Classification with Selected Features

This package is used to predict the discrete class labels based on a selected subset of high-dimensional features, such as expression levels of genes. The data are modeled with a hierarchical Bayesian models using heavy-tailed t distributions as priors. When a large number of features are available, one may like to select only a subset of features to use, typically those features strongly correlated with the response in training cases. Such a feature selection procedure is however invalid since the relationship between the response and the features has be exaggerated by feature selection. This package provides a way to avoid this bias and yield better-calibrated predictions for future cases when one uses F-statistic to select features.

Downloads: