Two-stage rule extraction method based on tree ensemble model for interpretable loan evaluation

Abstract

The tree ensemble model has been widely employed as a loan evaluation method in credit risk assessment due to its high accuracy and robustness. However, the tree ensemble model is complex and incomprehensible, which restricts its adoption for decision-making in loan evaluation. In this paper, we propose a novel rule extraction method for improving the ensemble model by balancing predictive performance and interpretability in two stages: a local rule extraction method followed by a global rule extraction method. The local method simplifies each rule by removing its redundant constraints, while the global method optimizes the complete rule set based on the multiobjective optimization method. An interpretable rule-based model is extracted from the tree ensemble model via the proposed method. Comparing the performance to six other methods on three loan evaluation datasets, the proposed method shows superior interpretability and realizes similar predictive performance to the tree ensemble model. For practical loan evaluation, the proposed method provides decision-makers with an interpretable rule-based model, which could replace the opaque tree ensemble model in high-stakes decision-making. In addition, the proposed method could facilitate decision-makers in explaining the tree ensemble model by analyzing the important and valuable rules that are extracted from the original opaque model.