Models and algorithms for safeguarded data-driven decision making

This thesis focuses on making smart decisions. Early works often equated smartness with the ability to do more with data (e.g., more accurate predictions, more profitable prescriptions), leading to ever progressing algorithms without safeguards. Safeguards are vital, however, to enforce fairness/ethics, interpretability, privacy, and robustness of decisions. A decision can only be truly smart if it is optimized while adhering to these values. With this perspective, this thesis focuses on the following key themes: (i) Developing safeguarded data-driven decision making models. I formulate safe-guarded data-driven decision making models to address concerns such as privacy and robustness. (ii) Deriving tractable algorithms to solve safeguarded problems. These safeguards often make the underlying computational tasks intractable; therefore, I derive efficient approximation schemes with rigorous performance guarantees for the optimization problems that arise in such settings. (iii) Analyzing safeguarding within multi-stage systems. Data-driven decision making involves multiple stages, typically starting with estimating the unseen truth from data and ending with optimizing decisions over the perceived reality. I aim to understand at which stage these safeguards should be imposed for the best outcomes.