Deep learning-based travel choice prediction with provable and adaptable fairness guarantees

Abstract

Deep Learning (DL) models offer substantial potential for travel choice predictions but are often plagued by algorithmic unfairness where disadvantaged population groups such as racial minorities and low-income populations often receive disproportionately worse prediction outcomes (e.g. accuracy) compared to their counterparts. Studies to address this issue in the transportation domain are relatively new and they fail to provide provable fairness guarantees and cannot address the diverse interpretations of fairness in practice. This study introduces a novel DL approach that provides provable fairness guarantees while being adaptable to various fairness standards. It embeds statistical hypothesis testing within a practical equality constraint to control disparities in prediction accuracy across different population groups, thus providing provable and adaptable fairness guarantees. This approach results in a threshold modification problem, formulated as a mixed-integer non-linear programming model that is proven to be NP-hard. To allow for efficient problem solving, theoretical properties of the threshold modification problem are investigated, enabling the decomposition of the original problem into smaller, more manageable subproblems. This decomposition provides insights into the problem's structure and enables the development of an efficient "Accuracy-First-Threshold-Second " algorithmic framework. Within this framework, an exact solution method is proposed to achieve optimal solutions, whereas a heuristic method, incorporating a sandwich algorithm and a bounded-enumeration algorithm, is designed to efficiently approximate near-optimal solutions. Extensive experiments demonstrate the computational performance of the proposed solution algorithms as well as the ability of the proposed fair DL approach to provide provable and adaptable fairness guarantees for travel choice predictions. This study offers a flexible and theoretically robust solution to fairness in travel choice prediction, with potential applications for enhancing equity in transportation systems.