Combining Machine Learning and Optimization Techniques for Decision-Making in Uncertain Scenarios

• Autores: Nuria Gómez Vargas
• Directores de la Tesis: Rafael Blanquero Bravo (dir. tes.) , Emilio Carrizosa Priego (dir. tes.)
• Lectura: En la Universidad de Sevilla ( España ) en 2025
• Idioma: inglés
• Número de páginas: 185
• Enlaces
  • Tesis en acceso abierto en: Idus
• Resumen

  • This PhD dissertation advances the integration of machine learning and optimization techniques for data-driven decision-making in uncertain environments. It addresses three fundamental sources of uncertainty in operations research: uncertain objectives in risk-neutral settings, where the goal is to optimize expected outcomes in the presence of unknown model parameters; uncertain constraints in risk-averse settings, where decisions must remain feasible across uncertain regions and perform well under worst-case realizations; and partially observed decision-making processes, where the underlying optimization problem must be inferred from observed data.

    The work focuses on contextual methods, in which decisions are tailored to auxiliary information—such as customer features or environmental variables—available at the time of decision-making. These methods enhance decision quality by adapting models to instance-specific characteristics, moving beyond fully specified models or simplistic uncertainty assumptions. A strong emphasis is also placed on interpretability, particularly through sparsity, which restricts the number of active features involved in decisions, resulting in simpler, more transparent, and more auditable models.

    The dissertation begins with a review of contextual decision-making under uncertainty and interpretability in optimization (Chapter 1). Chapter 2 introduces a profit-driven churn prediction model within a predict-and-optimize framework. By integrating the optimization objective into the learning stage and using individual customer lifetime values (CLVs), the model ensures that only the most valuable customers are targeted in retention campaigns. Unlike traditional strategies that rely on churn probabilities and average CLVs, this approach avoids aggregation bias. Optimized using stochastic gradient descent, the model shows superior performance across 13 real-world financial datasets.

    Chapter 3 presents a robust optimization framework for problems with temporally structured uncertainty. The model constructs context-aware uncertainty sets by conditioning vector autoregressive (VAR) processes on observed features, thereby capturing both temporal dependencies and contextual effects. Theoretical guarantees are provided, and numerical experiments confirm the method’s improved performance over static robust optimization approaches.

    Chapter 4 focuses on contextual inverse optimization in multiobjective settings, where the aim is to infer linear objective functions from observed decisions, assuming that decision-makers minimize a distance to an ideal point. The model supports multiple scalarizations (?1, ?2, ??, Ordered Weighted Averages) and includes structured sparsity constraints to promote interpretability. The resulting formulation is a Mixed-Integer Quadratic Program (MIQP), and empirical results show that the method successfully recovers sparse, interpretable models with strong generalization.

    Chapter 5 extends this line of work to a clustering-based inverse multiobjective optimization framework that uncovers both latent objectives and preference structures across multiple agents. Unlike conventional clustering methods that group by observed decisions, this approach clusters agents by their inferred trade-offs. The model minimizes optimality gaps under a MIQP formulation, includes a warm-start heuristic to improve scalability, and incorporates interpretability constraints. A real-world case study in diet recommendation highlights the framework’s ability to uncover robust and understandable decision patterns.

    Finally, Chapter 6 summarizes the main conclusions and outlines future research directions.