Generalized Priority-Aware Shapley Value
Brings cyclic, soft, and multi-criterion human preferences into LLM evaluation by modeling structured preferences as directed weighted priority graphs.
Research
Publications
My work develops statistically principled and computationally scalable methods for data valuation, explainable AI, LLM evaluation, and uncertainty quantification.
Google ScholarConference Papers
4 worksPriority-Aware Shapley Value
Introduces PASV, a Shapley-based attribution framework that incorporates precedence constraints and contributor-specific priority weights for more structure-faithful data valuation and feature attribution.
Faithful Group Shapley Value
Develops a robust and efficient method for group-level data valuation that is resistant to strategic group splitting and supported by theoretical guarantees.
Leave-One-Out Stable Conformal Prediction
A computationally efficient conformal prediction method that leverages leave-one-out stability to improve uncertainty quantification for black-box models.
* Equal contribution
Workshop Papers and Preprints
1 workFirst-Order Efficiency for Probabilistic Value Estimation via A Statistical Viewpoint
A statistical efficiency framework for estimating probabilistic values, including Shapley values and semivalues, with a new estimator guided by first-order mean squared error.
Journal Articles
1 workBayesian Hierarchical Time-Varying Distributed Lag Nonlinear Model: Applications in the Short-Term Association between Ambient Temperature and Daily Confirmed Cases of COVID-19
A Bayesian hierarchical model for studying time-varying nonlinear lagged associations, applied to temperature and COVID-19 confirmed cases.