ABSTRACT
Imputation of activity type from GPS traces has been considered of high importance in the domain of activity-travel behavior analysis. To increase the accuracy of activity type inference, recent studies have become increasingly dependent on the extensive use of personal location data, which are typically collected during the recruitment process or as part of auxiliary prompted recall sessions. Such data are however typically missing for increasingly more popular big datasets that tend to be aggregate in nature. To improve activity type imputation for such datasets, this paper proposes an enhanced approach for activity type identification from GPS measurements using recurrent profiles of individuals. It has been inspired by the fact that GPS-based data collection has shifted from single day to multi-day or multi-week data collection efforts, particularly for big data. The method derives from GPS panel data the frequency of activities by each predefined activity category. Results show that the temporal data play a much more important role than the spatial data in predicting activity types, supporting the relevance of the suggested approach for multi-day data. Considering the recurrent profiles of each activity, a model, which incorporates spatial and temporal variables and the frequency of activity locations, yields the best overall imputation accuracy of 67.4%.
