This study addresses the problem of managing traffic in urban transportation networks via path-based congestion pricing. Path-based congestion pricing policies consist of rewarding or tolling users for their path selections with the objective of mitigating system-wide congestion effects. The design of optimal path-based congestion pricing policies is notoriously difficult due to the large number of paths existing in transportation networks. Hence, both the problems of identifying candidate paths for pricing and that of computing optimal path-based congestion pricing policies are challenging. This study addresses these challenges by presenting novel statistical regression-powered optimization methods based on machine learning. We consider a bilevel optimization problem where a network planner aims to minimize congestion by providing path-based reward credits to commuters who are modeled as selfish agents minimizing a generalized cost function. We develop a statistical regression-powered heuristic approach to solve this path-based congestion pricing problem at scale. Our methodology integrates machine learning and optimization techniques to generate representative path sets and to model the relationship between congestion effects and path-based credit allocation. Sampling procedures and feature selection are used to synthesize a training data for a statistical regression model of congestion. Multiple supervised learning approaches are explored. The trained models are embedded in a surrogate optimization problem to determine path credits. An algorithm is designed to find feasible and efficient solutions to the bilevel optimization problem. Numerical experiments are conducted on small to large size networks. A comprehensive comparison is conducted between statistical regression-powered optimization methods, an exact branch-and-bound algorithm, and a model-based heuristic. The results demonstrate the efficiency and the computational scalability of the proposed statistical regression-powered optimization methods for solving problem instances based on large-scale networks. ... Read more