Bayesian Confirmation Measures in Rule-Based Classification

Abstract

With the rapid growth of available data, learning models are also gaining in sizes. As a result, end-users are often faced with classification results that are hard to understand. This problem also involves rule-based classifiers, which usually concentrate on predictive accuracy and produce too many rules for a human expert to interpret. In this paper, we tackle the problem of pruning rule classifiers while retaining their descriptive properties. For this purpose, we analyze the use of confirmation measures as representatives of interestingness measures designed to select rules with desirable descriptive properties. To perform the analysis, we put forward the CM-CAR algorithm, which uses interestingness measures during rule pruning. Experiments involving 20 datasets show that out of 12 analyzed confirmation measures \(c_1\), F, and Z are best for general-purpose rule pruning and sorting. An additional analysis comparing results on balanced/imbalanced and binary/multi-class problems highlights also N, S, and \(c_3\) as measures for sorting rules on binary imbalanced datasets. The obtained results can be used to devise new classifiers that optimize confirmation measures during model training.