Optimal Bayesian classification and its application to gene regulatory networks

A recently proposed Bayesian theory of classification can incorporate prior knowledge in the model to facilitate optimization and analysis for both classifier design and error estimation. Rather than rely on heuristic algorithms, this work is inspired by Wiener filtering in that it clearly states modeling assumptions and uses these to find optimal operators. The theory also gives rise to a sample-conditioned MSE, a new and useful tool for validating a proposed classifier. Herein, we summarize the theory and present an example classifying between normal and mutated gene regulatory networks based on the observed state of several genes. Partial prior knowledge is built into a discrete model, resulting in an optimal Bayesian classifier that can significantly outperform the popular discrete histogram rule.