A Practical Grafting Model Based Explainable AI for Predicting Corporate Financial Distress

Machine learning and deep learning are all part of artificial intelligence and have a great impact on marketing and consumers around the world. However, the deep learning algorithms developed from the neural network are normally regarded as a black box because their network structure and weights are unable to be interpreted by a human user. In general, customers in the banking industry have the rights to know why their applications have been rejected by the decisions made by black box algorithms. In this paper, a practical grafting method was proposed to combine the global and the local models into a hybrid model for explainable AI. Two decision tree-based models were used as the global models because their highly explainable ability could work as a skeleton or blueprint for the hybrid model. Another two models including the deep neural network and the k-nearest neighbor model were employed as the local models to improve accuracy and interpretability respectively. A financial distress prediction system was implemented to evaluate the performance of the hybrid model and the effectiveness of the proposed grafting method. The experiment results suggested the hybrid model based on the terminal node grafting might increase the accuracy and interpretability depending on the chosen local models.