Uncertainty is a pervasive challenge in decision-making and operations management. The era of data offers unprecedented opportunities to address this challenge while introducing new complexities. Data-driven decision-making, which enhances decision quality through advanced analytics, holistically integrates learning and optimization to mitigate uncertainty. In this thesis, we will discuss three projects about data-driven decision-making under uncertainty by integrating machine learning and optimization methodologies. First, we propose Fragility-aware Classification, introducing the Fragility Index—a risk-averse metric that quantifies the tail risk of confident misjudgments in classifiers. By employing the robust satisficing framework and developing tractable training methods, we enhance...[ Read more ]

Uncertainty is a pervasive challenge in decision-making and operations management. The era of data offers unprecedented opportunities to address this challenge while introducing new complexities. Data-driven decision-making, which enhances decision quality through advanced analytics, holistically integrates learning and optimization to mitigate uncertainty. In this thesis, we will discuss three projects about data-driven decision-making under uncertainty by integrating machine learning and optimization methodologies. First, we propose Fragility-aware Classification, introducing the Fragility Index—a risk-averse metric that quantifies the tail risk of confident misjudgments in classifiers. By employing the robust satisficing framework and developing tractable training methods, we enhance model robustness in safety-critical domains like medical diagnosis. Second, we address hotspot control in cloud computing by modeling computing resource allocation as a risk-averse bin packing problem. Leveraging time-series correlation analysis via piecewise aggregate approximation, we design allocation strategies to balance resource efficiency and risk mitigation in volatile environments. Third, we design a decision rule approach for two-stage mixed integer programming, unifying continuous and discrete uncertainty handling via a piecewise linear structure. We prove the optimality condition for the piecewise linear decision rules and show finite convergence of our iterative algorithm, enabling scalable solutions to stochastic programs with distributional ambiguity. Collectively, these contributions bridge between learning and optimization, offering novel frameworks to mitigate uncertainty, improve generalizability, and enhance decision-making.