On contextual inverse multiobjective problems
提出一个情境逆多目标优化框架，从情境-决策对中恢复多个目标函数和偏好，使用混合整数二次规划实现可解释的稀疏模型，并在真实数据上验证了泛化性能。
We introduce a novel framework for Contextual Inverse Multiobjective Optimization, where the goal is to recover multiple objective functions and the preferences among them from observed context-decision pairs. Assuming that each decision arises from minimizing a weighted distance to an ideal outcome, we develop a mathematical programming model that infers context-dependent linear objectives consistent with observed data. Our approach accommodates different norm-based scalarizations, including ℓ 1 , ℓ 2 , ℓ ∞ and Ordered Weighted Averages, and enforces interpretability through structured sparsity. By characterizing the conditions under which observed decisions are provably optimal, we formulate a tractable Mixed-Integer Quadratic Program that reveals how contextual features influence decision-making criteria. Numerical experiments on real-world data demonstrate the method’s ability to recover interpretable, sparse models while maintaining strong generalization performance.