Confirmation rule sets

Abstract

The concept of confirmation rule sets represents a frame-work for reliable decision making that combines two principles that are effective for increasing the predictive accuracy: consensus in an ensemble of classifiers and indecisive or probabilistic predictions in cases when reliable decisions are not possible. The confirmation rules concept uses a separate classifier set for every class of the domain. In this decision model different rules can be incorporated: either those obtained by applying one or more inductive learning algorithms or even rules representing human encoded expert domain knowledge. The only conditions for the inclusion of a rule into the confirmation rule set are its high predictive value and relative independence of other rules in the confirmation rule set. This paper introduces the concept of confirmation rule sets, together with an algorithm for selecting relatively independent rules from a set of all acceptable confirmation rules and an algorithm for the systematic construction of a set of confirmation rules.