Data Mining &
Decision Analytics

Inverse optimization is increasingly used to estimate unknown parameters in an optimization model based on decision data. However, when such "point estimates" are used to prescribe downstream decisions, the resulting decisions may be of low-quality and misaligned with human intuition, and thus less likely to be adopted. To tackle this challenge, we propose a novel decision recommendation pipeline that learns an uncertainty set for the unknown parameters and then solves a robust optimization model to prescribe new decisions. We show that the suggested decisions can achieve bounded optimality gaps, as evaluated using both the ground-truth parameters and human perceptions. Our method demonstrates strong empirical performance compared to the standard inverse optimization pipeline. Finally, we perform a case study where we apply this new pipeline to provide delivery route recommendations in Toronto, Canada. Our approach achieves a significantly higher delivery path adherence rate than current industry practices without compromising service quality. Moreover, our method provides a better trade-off between absolute and perceived decision quality than baselines under various realistic scenarios, including cases with model mis-specification and data scarcity.

Timothy Chan is the Associate Vice-President and Vice-Provost, Strategic Initiatives, and a Professor in the department of Mechanical and Industrial Engineering at the University of Toronto. His primary research interests are in operations research, optimization, and applied machine learning, with applications in healthcare, medicine, sustainability, and sports. He holds editorial roles in several journals including Operations Research, Management Science, and M&SOM. Recent honours include the 2026 JJ Berry Smith Doctoral Supervision Award from the University of Toronto, the 2026 CORS Award of Merit, the 2026 Innovative Application in Analytics Award from the INFORMS Analytics Society, the 2025 CORS Practice Prize, and finalist for the 2025 INFORMS Wagner Prize. Professor Chan received his B.Sc. in Applied Mathematics from UBC and his Ph.D. in Operations Research from MIT. Before coming to Toronto, he was an Associate in the Chicago office of McKinsey and Company.

Large language models (LLMs) are predominantly aligned with human preferences through reinforcement learning from human feedback (RLHF). In this talk, we explore the theoretical foundations of LLM alignment through the intertwined lenses of statistics and game theory. First, we show how the current formulation of RLHF induces a systematic bias we call preference collapse, and how this can be mitigated by introducing a tailored regularization term into the reward function. Next, we expose a fundamental bottleneck of reward-based alignment, demonstrating that cyclic human preferences cannot be faithfully represented by scalar reward models such as the Bradley-Terry model. More precisely, we establish that such cyclic inconsistencies give rise to a lower bound on the approximation error of any scalar reward fitting. Shifting to a game-theoretic perspective, we focus on Nash learning from human feedback and establish several social choice desiderata for this approach to alignment, including the preservation of preference diversity through the emergence of mixed strategies. Finally, we show that the zero-sum game approach generally cannot perfectly match a target preference distribution as a unique Nash equilibrium.

Weijie Su is a Professor of Statistics and Data Science at the Wharton School of the University of Pennsylvania, where he co-directs Penn Research in Machine Learning. His research examines the mathematical and statistical foundations of artificial intelligence, including large language models, privacy-preserving data analysis, high-dimensional inference, and optimization. He earned his Ph.D. in Statistics at Stanford University in 2016 and his bachelor's degree in Mathematics at Peking University in 2011. His honors include the 2026 COPSS Presidents' Award, the 2022 SIAM Early Career Prize in Data Science, the 2022 IMS Peter Gavin Hall Early Career Prize, a Sloan Research Fellowship, and an NSF CAREER Award. He is a Fellow of the Institute of Mathematical Statistics and a founding co-editor of Statistical Learning and Data Science.