Use of Sequential Bayes with Class Probability Trees

Abstract

Abstract For building classifiers from data, rule induction has established itself as an alternative to multivariate statistical approaches, including those of ‘neural’ computing. But modern rule-induction algorithms such as CART and C4 have not yet found a fully satisfactory way of discriminating logical from statistical forms of complexity in data. Systems of sequential Bayes rules offer a less ad hoc basis for combining probabilistic with rule-based approaches. In Evidencer, tree-structured rules are linked in a PROSPECTOR-like procedural hierarchy, and updated and processed according to a thresholding regime adapted from Wald’s sequential analysis (1947). The resulting formalism steers a course between the oversimplifications of PROSPECTOR and the complexity of full multi-level Bayes, while retaining precise error bounds on decisions.