Understanding Human Preferences for Counterfactual Explanations In Ai-Generated Loan Decisions

open_access
Open Access
Author:
Bislig, Jude
Area of Honors:
Data Sciences
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Amulya Yadav, Thesis Supervisor
John Yen, Thesis Honors Advisor
Keywords:
Counterfactual Explanations
Explainable Artificial Intelligence (XAI)
Human-Centered AI
Automated Loan Decision-Making
Algorithmic Fairness
Decision Tree Models
User Preference
Causal Reasoning in AI
Abstract:
In the modern day, the role of artificial intelligence continues to grow in critical decision-making processes, especially in loan approvals and finance. However, that does not mean it is without its own challenges. For one, there are ongoing challenges in making AI decisions fairer and more interpretable, as current decisions may lack transparency or may not give information to get a loan approved in the first place. Because of this, counterfactual explanations - hypothetical scenarios that illustrate the changes needed to achieve a different or more favorable outcome (Wachter et al., 2019) - can offer a way to explain loan rejections by presenting alternative scenarios or profiles where a given loan is approved. But what do humans generally prefer in terms of counterfactual explanations, and how do they perceive them? This study explores human preferences for these counterfactual explanations with two primary hypotheses: (1) humans will prefer a suboptimal but intuitive change over a more proximally optimal one, and (2) humans will prefer the counterfactual that avoids counterintuitive changes (e.g., reducing income) regardless of its evaluation metrics, such as proximity or validity. Findings from this research show that people prefer simpler and more human-like counterfactuals over more complex and precise ones. They also favored changes that improved their creditworthiness and profile overall. However, some people chose nonsensical counterfactuals over logical ones during research conducted, displaying how necessary it is that AI-generated decisions be human-centered. These results help to contribute more to the field of explainable AI (XAI) in finance and will aid in designing more user-friendly and transparent AI decision systems.