Machine learning for activity pattern detection

Abstract

This paper proposes a data fusion approach to automatically detect activity patterns in a GPS dataset based on travel diaries and correct misclassification errors. The Activity Patterns Detection consists of a Supervised Learning framework, thanks to which the activity purposes in the travel diaries are learned and then predicted in the GPS dataset. Furthermore, we deploy Unsupervised Learning to identify similar spatial and temporal activities in the GPS dataset and, based on travel diaries, to correct the misclassification errors. This work shows that, based on a few observations in the travel diaries and a set of features such as the resting time before the activity takes place, the number of occurrences of the same trip and the percentage of the trip made during daytime and the speed, it is possible to detect activities with an overall accuracy of 90%. Since the GPS dataset does not have information on the activity performed by the user, in reality, the aggregated results are validated based on the Kolmogorov-Smirnov test. The experiment shows that, with a confidence level of 99%, the majority of spatial and temporal feature distributions of activities in the travel diaries dataset are similar to those in the GPS dataset. Thanks to this approach, planners and transport operators can automatically obtain spatial and temporal patterns of frequent activities in urban areas.