The role of background knowledge in Bayesian classification

The development of Bayesian classifiers is frequently accomplished by means of algorithms that are highly data-driven. However, for many domains data-availability is scarce such that the resulting classifiers show poor performance. Even if performance is acceptable, Bayesian classifier structures are highly restricted and may therefore be unintelligable to the user. In this paper we address both issues. In the first part, we explore the trade-offs between classifiers constructed from clinical background knowledge and classifiers learned from a small clinical dataset. It is shown that the construction of classifiers from (partial) background knowledge is a feasible approach. In the second part, we introduce a construction algorithm that allows for a less restricted classifier structure, allowing easier clinical interpretation.