Explainable AI for dynamic ensemble models in high-stakes decision-making

Nishant Gadde 1, Avaneesh Mohapatra 2, *, Dheeraj Tallapragada 3, Karan Mody 4, Navnit Vijay 5 And Amar Gottumukhala 6

1 Jordan High School, Fulshear, Texas, United States.
2 Georgia Institute of Technology, Atlanta, Georgia, United States.
3 Dublin High School. Dublin, California, United States.
4 Irvington High School. Fremont, California, United States.
5 Acellus High School. Atlanta, Georgia, United States.
6 Heritage High School. Frisco, Texas, United States.
 
Review
International Journal of Science and Research Archive, 2024, 13(02), 1170–1176.
Article DOI: 10.30574/ijsra.2024.13.2.2091
 
Publication history: 
Received on 21 September 2024; revised on 10 November 2024; accepted on 13 November 2024
 
Abstract: 
AI is replacing human decisions with algorithmic ones in finance, healthcare, and justice, among other high-stake domains. We develop in this study the use of ensemble machine learning models in predictions related to credit default risks using two different datasets of credit records and application records. We achieve more robust predictions by using a combination of Random Forest, Gradient Boosting, and Decision Tree classifiers that use soft voting. Probably one of the most significant issues discussed in this research is class imbalance; defaults constitute only a really small fraction of cases compared to non-defaults in both datasets. In this respect, the Synthetic Minority Oversampling Technique was applied for balancing the classes by artificially creating synthetic samples of the minority class to have reasonably balanced training data. Ensemble model gave a respectable score of 0.88 on the ROC-AUC against the credit record dataset, where risk classification was dependent upon the 'STATUS' field. Applicants were divided into high-risk or low-risk candidates in case of an application record dataset, assuming income thresholds, and yielded an ideal score of 1.00 on the ROC-AUC. Precision-recall and ROC curves underlined the best class differentiation provided by models considering the imbalanced datasets. While these performances may have been quite strong for the ensemble models, the nature of the models is black-box, which can be an issue in financial domains where transparency is very much expected along with accountability. To this end, we suggest the inclusions of XAI techniques in future work: dynamically and in real time, explanations of model decisions. It will be certain that stakeholders have a belief not only in the accuracy of AI-driven decisions but also in the rationale behind them-developing trust and enhancing interpretability for credit risk predictions.
 
Keywords: 
AI; High-stakes decision-making; Human decisions; finance; Healthcare
 
Full text article in PDF: 
 
Copyright information: 

Copyright © 2024 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0