A mode-constrained mixture model for route choice inference

Cities increasingly rely on bike-sharing systems to promote sustainable mobility. These systems generate massive but low-granularity data. Unlike GPS datasets, trip records typically provide only the origin, destination and total duration, making it difficult to infer route structure or behavioral heterogeneity. Despite this limited granularity, some studies have shown that trip-duration distributions are usually well represented by a log-normal mixture model. The first component of the mixture model is believed to provide a faithful representation of the main route travel times, with a strong linear relationship existing between its mode and the fastest route provided by OpenStreetMap. We exploit this relationship to propose a more robust analysis of the travel time distribution via a mixture model with a mode-constrained first component. This approach is validated on the Helsinki-Espoo data which includes both trip durations and distances, the latter allowing to clearly identify the various trip behavior groups of the mixture model due to its lesser variability. The estimated group proportions are then confronted to those estimated from the durations as a way to validate the method. This transferable method offers urban planners a reliable tool to infer route choice without GPS data, supporting the design of cycling infrastructure and the modeling of travel demand.