Abstract
Predictive modelling is one of the most important data mining tasks, where data mining models are trained on data with ground truth information and then applied to previously unseen data to predict the ground truth of a target variable. Ensemble models are often used for predictive modelling, since ensemble models tend to improve accuracy compared with standalone classification models. Although ensemble models are very accurate, they are opaque and predictions derived from these models are difficult to interpret by human analysts. However, explainability of classification models is needed in many critical applications such as stock market analysis, credit risk evaluation, intrusion detection, etc. A recent development of the authors of this paper is ReG-Rules, an ensemble learner that aims to extract a classification (prediction) committee, which comprises the first rule from each base classifier that fired. The rules are interpretable by humans, thus ReG-Rules is a step towards explainable ensemble classification. Since there is a set of matching rules presented to the human analyst for each prediction, there are still numerous rules that need to be considered for explaining the model to the human analyst. This paper introduces an extension of ReG-Rules termed Consolidated Rules Construction (CRC). CRC merges individual base classification models into a single rule set, that is then applied for each prediction. Only one rule is presented to the human analyst per prediction. Therefore, CRC is more explainable than ReG-Rules. Empirical evaluation also shows that CRC is competitive with ReG-Rules with respect to various performance measures.