Restoring coverage to the Bayesian false discovery rate control procedure

Principal among knowledge discovery tasks is recognition of insightful patterns or features from data that can inform otherwise challenging decisions. For the costly future decisions, there is little room for error. Features must provide substantial evidence to be robust for classification and dependable for important decisions. Here we seek statistical evidence for feature selection, that feature signals are of sufficient magnitude and frequency to be generalizable for classification. The Bayesian false discovery rate (bFDR) error control procedure is powerfully suited for this task. In realistic situations often encountered in practice, the bFDR procedure is biased, yielding a greater than desired FDR. In other less typical cases, the FDR is less than desired. We investigate the sources of bias in the bFDR procedure, and predict the direction of bias. A new algorithm has been developed to recover the bias in the bFDR control procedure. In simulation and real data mining examples, the new bFDR control algorithm shows promise. The strengths and limitations of the new approach are presented with examples and discussed.