Semi-automatic imputation of long-term activity-travel diaries using GPS traces: personal versus aggregate histories

The new generation of dynamic models of activity-travel demand requires multi-day or multiweek activity-travel data. A combination of modern GPS technology and a prompted recall instrument may be a powerful tool to reduce respondent burden and collect such multi-week data of travel behavior. The authors have developed such a system, called TraceAnnotator which was designed for automatic imputation of various facets of activity- travel patterns form GPS tracers. The core of the system developed is a Bayesian belief network that classifies the outcome variables of interest, using a network of input variables. This means that the interpretation of the GPS traces of any new respondent is based on the aggregate conditional probability tables that the system learned on the basis of the previously processes respondents/cases. However, in case of multi-day data collection, the history of every respondent is also collected. This implies that learning can be based on the continuously updated conditional probability tables, aggregated across respondents, or on the personal histories of respondents or on a combination of both. This paper will discuss the results of these alternative approaches to impute transport modes and activity types for multi-week activity-travel diary data using GPS technology.