OpFlow Improves Robustness in OD Flow Prediction

Origin-destination (OD) flow prediction is a cornerstone of urban analytics, but current deep learning models often struggle with distribution shifts because they are trained on raw counts. This approach conflates transferable choice mechanisms with environment-specific factors, making models brittle when conditions change. The core issue is that raw OD counts mix two distinct elements: the total demand generated by an origin and how that demand is distributed among various destinations. This research proposes OpFlow, a novel framework designed to overcome these limitations by focusing on the transferable aspect: the "exposure-to-choice law" that maps spatial conditions to relative destination preferences. OpFlow employs a mechanism-constrained approach, learning row-centered choice potentials and then reconstructing flows by combining these induced allocations with a separately calibrated origin scale. This design allows spatial exposures and allocations to vary under distribution shifts, while the conditional mapping from exposure states to relative choice potentials remains robust. Theoretical analysis characterizes the identifiable row-centered potential, showing that classical spatial interaction laws are specific cases of restricted log-potentials. Both synthetic experiments and a real-world application demonstrate that OpFlow significantly enhances robustness when faced with environmental shifts.