Graph Supported Mode Detection within Mobile Phone Data Trajectories

Mobile phone data (MPD) has been used in various studies to analyze human travel behavior in time and space. While the pure number of available trips is promising, data quality and the information density of each individual trajectory is relatively low. Travel mode detection of MPD-trajectories is a challenging task, since geographic location data is noisy and the events within a trajectory are irregular (event-based instead of time discrete). In this paper, we present a method to identify travel modes typically observed in urban areas, such as walking, bicycle, tram, bus, and car. The method requires concise network graphs for each mode. Annotated GPS trajectories were collected from volunteers as ground truth to train and validate various machine learning algorithms. The cleaned trajectories of the MPD are segmented into individual trips, which are mapped on the network graphs using a map-matching algorithm. Various features, such as trip distance and travel speed, were analyzed to identify the most suitable features for classifying the available modes with a random forest (RF) and a support-vector machine algorithm. With the RF algorithm, about 80% of all trips were associated to the correct mode. Since the total dataset was only comprised of about 600 trips, which then needed to be split into an evaluation dataset and a training dataset, we suspect the accuracy of the method will increase with more data. Based on the results, this work presents a proof of concept for determining the travel mode of MPD-trajectories.