Over the past couple of decades, there has been an exponential increase in the collection of large-scale
GPS data from household/personal travel surveys all over the world. A range of algorithms, which vary
from specific rules to advanced machine learning methods, have been applied to extract travel modes
from raw GPS data collected by smartphone-based travel surveys. However, most of the methods applied
neither describe the interaction between features influencing the travel mode decision nor effectively
deal with the ambiguity inherently incorporated in these features. This paper identifies travel modes with
a Bayesian network, whose structure is established based on a K2 algorithm and corresponding conditional
probability tables are estimated with maximum likelihood methods. Five representative travel
modes – walk, bike, e-bike, bus and car – are distinguished using the resulting Bayesian network.
Additionally, the low speed rate and the average heading change are introduced to reduce uncertainties
between bike and e-bike segments and between bus and car segments. The derived travel modes are then
compared with those retrieved in the prompted recall survey by telephones. Consequently, more than
86% of segments have the travel mode correctly identified for each travel mode, with over 97% of walk
segments being properly flagged. Results from the study demonstrate that GPS travel surveys provide
an opportunity to supplement traditional travel surveys.