DS Ph.D. Qualifier Presentation | Peter M. VanNostrand | December 8th @ 10:00AM | Salisbury Labs 105
10:00 a.m. to 11:00 a.m.
United States
Ph.D. Qualifier Presentation
Peter M. VanNostrand
Friday, December 8th, 2023 | 10:00AM - 11:00AM
Location: Salisbury Labs Room 105
Committee:
Prof. Elke A. Rundensteiner, Advisor, WPI
Prof. Andrew C. Trapp, Qualifier Co-advisor, WPI
Prof. Xiaozhong Liu, Qualifier Co-advisor, WPI
Title: What Should I Do? Examining Lay User Understandings of Counterfactual Explanations of AI Decisions for Actionable Recourse
Abstract:
Machine learning systems are increasingly deployed to automate decision making in domains such as hiring and credit approval where negative outcomes can have substantial ramifications for decision subjects. There has been sizeable research in developing Explainable AI tools to provide algorithmic transparency. Of particular focus has been the generation of counterfactual explanations which aspire to empower decision subjects to seek actionable recourse by describing alternations that would transform a negative outcome to a positive one. While these works have achieved promising theoretical results, there has been limited empirical evaluation of if lay users understand and can work with counterfactual explanations and of how best to present counterfactual information. To bridge this gap, we conduct an online between-subjects crowd-sourced user study (N=252) to examine lay user understandings of counterfactual explanations considering two main factors: explanation style and explanation presentation. We find that lay users have a significantly higher objective and subjective understanding of the AI decision-making system as well as an increased user confidence when using state-of-the-art region-style counterfactual explanations as compared to traditional point-style counterfactuals. We also find that the presentation style of the explanation has a significant moderating effect on user confidence but not on understanding. Lastly, visual presentations of explanations lead to faster response times than natural language presentations.