Publications
- 2026
Penalty-Based First-Order Methods for Bilevel Optimization with Minimax and Constrained Lower-Level Problems
Yiyang Shen*, Yutian He*, Weiran Wang, Qihang Lin
Accepted at the 40th Annual Conference on Neural Information Processing Systems (NeurIPS 2026)
* Equal contribution; author ordering determined by a coin flip.
Preprint ↗ - 2026
Enforcing Fair Predicted Scores on Intervals of Percentiles by Difference-of-Convex Constraints
Yutian He, Yankun Huang, Yao Yao, Qihang Lin
Proceedings of the 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026), PMLR 300:2269–2277
Read paper ↗ Poster (PDF) ↗ - 2022
Optimal layout of stacked graph for visualizing multidimensional financial time series data
Yutian He, Hongjun Li
Information Visualization, 21(1), 63–73
Journal article ↗
Working Papers
Distributionally Fair Bilevel Optimization
Yutian He, Beste Basciftci, Qihang Lin
We study distributional fairness in stochastic hierarchical decision making. A leader chooses a decision before uncertainty is observed, a follower subsequently optimizes a scenario-dependent response, and the resulting decisions induce random utilities for multiple communities. Read moreRead less
We measure unfairness by the maximum pairwise squared Wasserstein distance between the communities' utility distributions and minimize this measure subject to an upper bound on the leader's expected cost. This formulation captures fairness beyond expected outcomes while recognizing that the leader can influence, but cannot directly choose, the follower's actions. Using sample average approximation, we develop two complementary solution approaches. First, we derive an exact mixed-integer nonlinear reformulation based on empirical quantiles and lower-level optimality conditions. Second, under difference-of-convex (DC) representability of the model functions and lower-level value function, we obtain an equivalent DC-constrained formulation and propose an inexact DC algorithm for computing a nearly ε-KKT point with an explicit iteration-complexity guarantee. We also show how the value functions and their subgradients can be computed efficiently in two motivating applications: capacity planning and a newsvendor pricing game. Computational experiments demonstrate that the continuous method finds high-quality solutions, often within a few minutes and with small gaps relative to available lower bounds, whereas the exact Gurobi formulation frequently reaches a four-hour time limit. The results show that distributional fairness can be incorporated into large stochastic leader--follower models without sacrificing computational practicality.
Decision Dependent Distributionally Robust Optimization for Facility Location Problem with Customer Behavior Considerations
Yutian He, Beste Basciftci, Xian Yu
Facility location problems are often formulated as two-stage stochastic optimization models, where facility decisions are made in the first stage before uncertainty is realized, and second-stage recourse decisions are used to respond to uncertain demand or customer behavior. Read moreRead less
However, in many applications, customer choices are not independent of the first-stage decisions: the selected facility locations can directly influence consumer preferences and subsequent behavior. This work addresses such a facility location problem with customer behavior considerations. We develop a decision-dependent distributionally robust optimization model that captures customer behavior as second-stage recourse. Specifically, we utilize a ranking-based choice model informed by machine learning to represent consumer preferences, allowing customer choices to depend on the selected facility locations. We formulate the resulting problem as a decision-dependent distributionally robust mixed-integer bilevel model, derive a single-level reformulation, and investigate the model under different forms of ambiguity sets. To support the data-driven component, we use a feedforward neural network to learn customer rankings from observed data. We further develop an enhanced column-and-constraint generation (CCG) algorithm to improve computational efficiency. Using EV charging station planning as a representative application, we demonstrate the effectiveness of the proposed modeling and solution approach.