Behavioral cycling profiles and their potential as estimators for cycling motives

Cycling behavioral research is increasingly conducted by means of GPS data. The presence of\nthese data-sets allow for large-scale investigation of complicated travel behavioral aspects such as\ncycling motives and enables one to enrich raw GPS data-sets with these attributes based on\ncontextual information. Currently, both the differences in cycling behavioral between cyclists with\ndifferent motives and the extent up to which these differences can be used to estimate cycling\nmotives for raw GPS tracks have received little attention. Even though more insights on these\ntopics can provide useful insights for policymakers and can stimulate travel behavior research by\nenabling others to enhance their GPS tracks with more accurate cycling motive attribute data.\nThis research tries to tackle both these problems by establishing cycling behavioral profiles based\non trip, route and origin-destination behavioral characteristics and subsequently using the\ndifferences in these profiles to estimate cycling motives by means of machine learning. In addition\nto that, multiple machine learning algorithms are assessed to determine the most suitable The\nresults show that there are significant differences in cycling behavioral profiles between motives.\nTrip, route and origin-destination behavioral characteristics all outperform a standard model for\nestimating cycling motives, with a combined model including all behavioral characteristics scoring\nhighest (74.0% accuracy versus 51.4% standard model accuracy). Furthermore the results indicate\nthat Random Forest and Gradient Boosting are among the most suitable algorithms for this\npurpose. Finally, recommendations and potential improvements are provided for future research on\ncycling behavior and motive estimation.