Research
My research studies Explainable AI (XAI), Interpretable Machine Learning, and Human–AI Decision Making, focusing on how to generate AI explanations that are credible and risk-sensitive, how to explain predictions when input data are missing, and how different sources of uncertainty shape human trust in AI systems.
Conference Papers
Counterfactual Explanations for AI Model Predictions with Missing Inputs
Most existing explainable AI methods assume complete input data and, when values are missing, fall back on simple imputation before generating explanations. This paper studies how to produce counterfactual explanations for undesired predictions when input data are incomplete. We propose a method that both recommends improvements to observed features and identifies which missing values would most change the prediction if completed, using a SHAP-based feature importance measure to guide the selection of counterfactuals. Rather than defaulting to mean imputation, our approach encourages disclosure of the missing information most useful for reaching a desired outcome. An empirical evaluation on real-world data shows the method produces more informative and actionable counterfactual explanations than baselines that ignore missing inputs.
More Than Confidence Scores: Do Different Types of Uncertainty Matter for Trust in AI?
Communicating uncertainty is essential for effective human–AI collaboration because it shapes user trust and reduces both overreliance and unwarranted aversion. This study examines how aleatoric (data-related) and epistemic (model-related) uncertainty differentially affect trust and decision behavior. In a two-stage controlled experiment (N = 240) using a high-stakes health insurance cost prediction task, both forms of uncertainty significantly reduce behavioral reliance on AI predictions relative to a no-uncertainty baseline, with data-driven (aleatoric) uncertainty producing a stronger reduction than model-driven (epistemic) uncertainty. Transparent communication of uncertainty modestly improves perceived satisfaction and usefulness for both types. The findings show that the source of uncertainty is not interchangeable in its effects on trust, informing the design of more effective uncertainty-communication strategies for high-stakes AI applications.
Risk-Sensitive Counterfactual Explanations for AI Model Predictions
This paper presents a novel approach to risk-sensitive counterfactual explanations, improving AI prediction interpretability by incorporating the risk associated with a prediction. Unlike traditional counterfactuals that neglect the risk of misleading explanations, our method delivers robust counterfactuals for high-risk scenarios and proximal counterfactuals for lower-risk cases, ensuring actionable and risk-aware insights. Evaluated on two real-world datasets, our approach outperforms state-of-the-art baselines on average risk, sparsity, and Local Outlier Factor (LOF) metrics.
Counterfactual Explanations for Incorrect Predictions Made by AI Models
This paper proposes a novel method to generate counterfactual explanations for misclassified instances from black-box models, using a decision tree fitted to classification results and trust scores to correct errors and tailor explanations for beneficial-to-adverse and adverse-to-beneficial misclassifications. Evaluated on two real-world datasets against three state-of-the-art XAI methods, it improves accuracy by correcting misclassifications and provides more plausible, sparse explanations.