SHAP and LIME: An Evaluation of Discriminative Power in Credit Risk

• DOI: 10.3389/frai.2021.752558
• Publication Date: 2021-09-17T05:48:36Z
• Authors: Alex Gramegna; Paolo Giudici
• Keywords: Discriminative model; Boosting; Gradient boosting; Predictive power

In credit risk estimation, the most important element is obtaining a probability of default as close possible to effective risk. This effort quickly prompted new, powerful algorithms that reach far higher accuracy, but at cost losing intelligibility, such Gradient Boosting or ensemble methods. These models are usually referred “black-boxes”, implying you know inputs and output, there little way understand what going on under hood. As response that, we have seen several different Explainable AI flourish in recent years, with aim letting user see why black-box gave certain output. this context, evaluate two very popular eXplainable (XAI) their ability discriminate observations into groups, through application both unsupervised predictive modeling weights these XAI assign features locally. The evaluation carried out real Small Medium Enterprises data, obtained from official italian repositories, may form basis for employment post-processing extraction.