Interpretable Multiclass Models for Corporate Credit Rating Capable of Expressing Doubt

Corporate credit rating is a process to classify commercial enterprises based on their creditworthiness. Machine learning algorithms can construct classification models, but in general they do not tend to be 100% accurate. Since they can be used as decision support for experts, interpretable models are desirable. Unfortunately, interpretable models are provided by only few machine learners. Furthermore, credit rating often is a multiclass problem with more than two rating classes. Due to this fact, multiclass classification is often achieved via meta-algorithms using multiple binary learners. However, most state-of-the-art meta-algorithms destroy the interpretability of binary models. In this study, we present Thresholder, a binary interpretable threshold-based disjunctive normal form (DNF) learning algorithm in addition to modifications of popular multiclass meta-algorithms which maintain the interpretability of our binary classifier. Furthermore, we present an approach to express doubt in the decision of our model. Performance and model size are compared with other interpretable approaches for learning DNFs (RIPPER) and decision trees (C4.5) as well as non-interpretable models like random forests, artificial neural networks, and support vector machines. We evaluate their performances on three real-life data sets divided into three rating classes. In this case study all threshold-based and interpretable models perform equally well and significantly better than other methods. Our new Thresholder algorithm builds the smallest models while its performance is as good as the best methods of our case study. Furthermore, Thresholder marks many potential misclassifications in advance with a doubt label without increasing the classification error.