北大经院工作坊第1313场
Efficient and Interpretable Transformer for Counterfactual Fairness
风险、保险与不确定性经济学工作坊
主讲人：全智雨（伊利诺伊大学厄巴纳香槟分校助理教授）
主持人：
（北大经院）陈凯
（人大财金）陈泽、胡文涛
（清华经管）冯润桓
参与老师：
（北大经院）郑伟、贾若
（人大财金）魏丽、何林
时间：2026年6月24日（周三）10：00-11：30
线上形式：腾讯会议
会议号：205 870 222
线下地点：北京大学经济学院303会议室
主讲人简介：
Zhiyu (Frank) Quan is an Assistant Professor at the Department of Actuarial and Risk Management Sciences of the University of Illinois Urbana-Champaign, a Brad and Karen Smith Professorial Scholar of the College of Liberal Arts and Sciences, and a Finance and Insurance Sector Lead for Discovery Partners Institute. He holds a Ph.D. in Actuarial Science from the University of Connecticut. Before joining Illinois, he worked in a cutting-edge Insurtech company as a R \& D data scientist developing data-driven solutions for major insurance companies. He has a broad spectrum of research interests in data science applications in insurance such as tree-based models, natural language processing, GenAI, and applies his actuarial expertise to build predictive models for claim research, rate making, etc. His research agenda is driven by real-life data and is inspired by collaborations with InsurTech and insurance companies. Besides, he is a director of the Illinois Risk Lab, which facilitates research activities that integrate academic training with practical problem-solving in real business settings. He has received 2021 Arnold O. Beckman Research Award, 2025 Bob Alting von Geusau Prize, and has been awarded by the Society of Actuaries Research Institute and Casualty Actuarial Society.
摘要：
The growing reliance of machine learning models in high-stakes, highly regulated domains such as finance and insurance has created a growing tension between predictive performance, interpretability, and regulatory fairness requirements. In these settings, models are expected not only to deliver reliable predictions but also to provide transparent decision rationales and comply with strict fairness requirements. Existing fairness-aware learning methods for tabular data, however, often focus primarily on group-level fairness metrics or depend on explicit and structural causal model assumptions that are challenging to validate in practice. Meanwhile, attention-based transformers offer powerful mechanisms for modeling complex data relationships as demonstrated in various language tasks, yet their attention mechanisms alone do not ensure counterfactually fair predictions, even when combined with fairness-aware techniques. To address these limitations, we propose the Feature Correlation Transformer (FCorrTransformer), an attention-light architecture tailored for tabular data. In this design, the attention matrix admits a direct statistical interpretation as pairwise feature dependencies, enhancing both interpretability and efficiency. Leveraging this structure, we introduce Counterfactual Attention Regularization (CAR), a framework that enforces group-invariant fair representations of sensitive features at the attention level, promoting counterfactually fair predictions without relying on explicit causal assumptions. Empirical evaluations on imbalanced classification and regression benchmarks demonstrate that FCorrTransformer combined with CAR achieves strong counterfactual fairness while maintaining competitive predictive performance and substantially reducing model complexity compared with standard transformer-based baselines. Overall, this work bridges a critical gap between fairness theory and machine learning models, offering a practical framework for responsible AI in regulatory-sensitive domains.
供稿：科研与博士后办公室
美编：初夏
责编：度量、雨禾、雨田
