Origin--destination (OD) matrices are essential for forecasting and capacity planning in transportation and communication networks, yet they are not directly observable and must be inferred from limited measurements. Classical gravity and power-law models capture broad distance effects but lack theoretical grounding and enough practical precision, while data-driven approaches require large training sets and are brittle under architectural change. We introduce a probabilistic framework that derives OD flows from individual interactions. The model separates end-to-end traffic into two observable components: how often individuals in two locations interact (their attractiveness) and the average traffic carried by each interaction. This decomposition provides a physically interpretable, data-fusable structure that remains valid regardless of network architecture. To evaluate the model, we focus on the empirically dominant case in which interaction probabilities decay with distance according to a power law. From this, we derive an origin-specific attractiveness measure in which the effective distance is scaled by the origin’s population density and calibrated independently for each origin. Applied to large-scale county–pair social-interaction data, the resulting origin-specific fits substantially outperform classical, globally parameterized power-law models and recover exponents consistent with independent empirical estimates. These findings reinforce the core implications of the model: OD flows scale linearly with population, and the dominant source of heterogeneity stems from origin-specific interaction behavior rather than nonlinear population effects. This provides a practical pathway for robust, high-precision, end-to-end OD estimation, especially given the widespread availability of large-scale statistical records.