Learning transportation mode from raw gps data for geographic applications on the web

Geographic information has spawned many novel Web applications where global positioning system (GPS) plays important roles in bridging the applications and end users. Learning knowledge from users' raw GPS data can provide rich context information for both geographic and mobile applications. However, so far, raw GPS data are still used directly without much understanding. In this paper, an approach based on supervised learning is proposed to automatically infer transportation mode from raw GPS data. The transportation mode, such as walking, driving, etc., implied in a user's GPS data can provide us valuable knowledge to understand the user. It also enables context-aware computing based on user's present transportation mode and design of an innovative user interface for Web users. Our approach consists of three parts: a change point-based segmentation method, an inference model and a post-processing algorithm based on conditional probability. The change point-based segmentation method was compared with two baselines including uniform duration based and uniform length based methods. Meanwhile, four different inference models including Decision Tree, Bayesian Net, Support Vector Machine (SVM) and Conditional Random Field (CRF) are studied in the experiments. We evaluated the approach using the GPS data collected by 45 users over six months period. As a result, beyond other two segmentation methods, the change point based method achieved a higher degree of accuracy in predicting transportation modes and detecting transitions between them. Decision Tree outperformed other inference models over the change point based segmentation method.