Grant: Discovery Grants Program - Individual

Machine learning and optimization are two key technologies that drive the development of analytics in today's industries. The increasing availability of large datasets has motivated the design of new machine learning and optimization models that can better exploit the data. Central to the design of these models are the principles of data-drivenness and robustness. The former requires a model to effectively capture the information contained in observed data, whereas the latter requires a model to offer protection against sampling errors or unforeseen data. An emerging framework for implementing these principles is data-driven distributionally robust optimization (DD-DRO), which substitutes the empirical data (distribution) with a set of distributions that are "close" to the empirical distribution and finds a solution that performs well against any distribution from the set. This framework has its intuitive appeal, but it provides very limited guidance regarding how to best quantify the "closeness" to the empirical distribution and suffers from a hard trade-off between data-drivenness and robustness when it comes to choosing a distance metric to quantify the "closeness". The goal of my Discovery program is to establish a new DD-DRO framework that can resolve, or at least lessen, the trade-off between data-drivenness and robustness arising from the existing framework. My objectives are: 1) establishing a new DD-DRO framework by generalizing existing models and developing new distance metrics suited for the simultaneous pursuit of data-drivenness and robustness, 2) applying the new framework to develop new models for machine learning, inverse optimization, and fairness (equity) optimization, 3) validating the value of the new ML and optimization models in the context of loan portfolio management, using real-life data, where fairness to minority groups is of priority Resolving this trade-off will have a groundbreaking impact on the development of decision support systems (DSS), particularly on supporting resource decisions such as loan portfolio management, which requires justification in terms of alignment with collected data and, at the same time, protection against unforeseen uncertainties. Analytics-powered DSS, while increasingly applied in portfolio management, have been criticized for their untrustworthiness and biases. Our framework, built upon the principles of data-drivenness, robustness, and fairness, will enable the building of portfolios that are more efficient, resilient, and equitable. This in turn provides financial stability to the economy and financial resources to the disadvantaged. My program not only offers innovation with global impact but also equips students with both analytics and business skills that are highly valued in industries such as finance and other data-driven sectors.