Rule Mining Based on Nonmonotonic Rule Layers and Its Application to Medicine

This paper proposes a new framework for rule induction methods based on rule layers constrained by inequalities of accuracy and coverage. Rule layers consists of (regular) rule layer, in subrule layer out subrule layer, and non-rule layer. Since the rules can be classified into one of the layers by the inequalities, updates of probabilistic rules are equivalent to their movement between layers. Rule induction methods are combined with classification of elementary relations (i.e., one attribute-value pair) into four layers. The classification statistics reflects the characteristics of data and rule stabilities. The proposed method was evaluated on datasets regarding headaches and meningitis, and the results show that the proposed method not only outperforms the conventional method but also captures the characteristics of applied data.