Data-Side Model Reduction For Large-Scale Supply Chain Optimization: Aggregation, Sparsification, And Preprocessing Trade-Offs Between Solve Time And Solution Fidelity

Abstract

When an optimization model reaches continental scale, the solver stops being the constraint and the data becomes it. A solver cannot compensate for bad inputs; it will faithfully optimize the wrong problem. More consequentially, what is fed to the solver defines the solution space: dominated lanes, redundant paths, and low-volume noise expand the model without adding decision value, inflating solve time while diluting solution quality. Data-side reduction is therefore problem definition rather than preprocessing. This paper describes a modular, configuration-driven reduction pipeline that replaced a monolithic input-generation system feeding connectivity and timing optimization models for a continental logistics network spanning more than twelve hundred sites. Six reduction techniques are applied in a deliberate cascading order, because structural reductions propagate downstream and must precede temporal and flow aggregation. The paper states where reduction must stop, defining three hard boundaries: feasibility, service-level fidelity, and volume-redistribution accuracy. It reports measured outcomes. Network nodes fell from more than twenty thousand to approximately two thousand, origin-destination pairs from tens of millions to hundreds of thousands, and temporal records by roughly seventy percent. Connectivity and timing model solve time fell from more than thirty hours to more than twelve. Percentage to optimality improved from seventy-five to ninety. Four reductions that looked reasonable and were unsafe are reported in full, each having produced feasible-looking output that was wrong, together with the architectural response each motivated.