XAI in the model exploration process for credit scoring
Nowadays, interpretability became a hot topic, which leads to the increasing awareness of the need for including eXplainable Artificial Intelligence (XAI) in the life cycle of models. Hence, the main need is no longer to convince people of the usefulness of explainability but to establish domain-specific procedures for XAI-based model development.
In this post, I will show how to use XAI in credit scoring.
Presented concepts come from the paper Transparency, auditability, and explainability of machine learning models in credit scoring that is joint work with Michael Bücker, Gero Szepannek, and Przemysław Biecek.
Logistic regression is currently considered the gold standard methodology in the credit scoring industry. The reason for that is the simplicity and interpretability that give customers easy explanations of individual decisions and make it easier for auditors to validate such models.
It might seem that complex algorithms with thousands of coefficients are on the opposite side. They are considered opaque and difficult to understand. However, including XAI methods allows us to understand some aspects of such models and make them more interpretable. In credit scoring, the ability to understand the model is not only desirable but also regulated. Therefore, the use of XAI methods is the only gateway to utilizing more complex, and thus potentially more effective, algorithms.
The procedure of developing and maintaining a credit scoring model is complex and involves many participants. The core elements of such procedure include the activities of the stakeholders across the model’s life, the needs of each stakeholder, and the XAI methods appropriate for each need. Below, I show the systematic exploration process focused on Transparency, Auditability, and eXplainability for Credit Scoring models (TAX4CS).
Who? The stakeholders that participate in the TAX4CS are, among others, the model developers, auditors, credit officers, and customers.
When? The stakeholders' involvement takes place at various stages. Model developers are mostly involved at the beginning of the model life cycle. The auditors' role begins during the validation process and then is carried out periodically to monitor the effectiveness of a model. During the production stage credit officers and customers continuously utilize the model’s predictions.
What? The Stakeholders have diverse needs. Model developers should be able to establish which model is the most effective. The auditors need to manage checks required by regulators. Credit officers and customers should be able to understand the key features behind the credit decision.
How? XAI methods offer the opportunity to answer the needs of different stakeholders. Model developers and auditors can use global-level model explanations that place greater emphasis on the suitability of the model instead of interpretability only, for example, performance measures and variable importance measures. Local-level model explanations are more suitable for credit officers and customers. Methods, such as Break Down or Ceteris Paribus allow for the analysis of a single observation and therefore help to understand the factors behind credit decisions.
If you are interested in this topic, I encourage you to read our paper that provides a detailed overview of XAI techniques for credit scoring with a real-world case study.
Michael Bücker, Gero Szepannek, Alicja Gosiewska & Przemyslaw Biecek (2021) Transparency, auditability, and explainability of machine learning models in credit scoring, Journal of the Operational Research Society, DOI: 10.1080/01605682.2021.1922098.
