Stop Purpose Classification from GPS Data of Commercial Vehicle Fleets

Extracting stop purpose information from raw GPS data is a crucial task in most location-aware applications. With the continuous growth of GPS data collected from mobile devices, this task is becoming more and more interesting; a lot of recent research has focused on pedestrians (mobile phones) data, while the commercial vehicles sector is almost unexplored. In this paper we target the problem of stop identification and classification from vehicle GPS data, using a large and heterogeneous dataset of commercial fleets from diverse industries. Our aim is to classify stops by purpose in two categories: work related and non-work related. Our dataset consists of more than 700k stops, 160k of which are work stops. For each stop, we compute a set of 100 different features, which can be grouped in 4 main categories: stop-wise features, points of interest features, stop cluster features, and sequential features. By choosing Random Forests as classification model, we are able to assess the relative importance of each of the features in the four sets. Experimental results show that our method significantly outperforms the state of the art models for stop purpose classification in the context of commercial vehicles. The feature ranking highlights the importance, for classifying a stop, of both its duration and the duration of other stops in the same location