SSMDL: Semi-supervised Multi-task Deep Learning for Transportation Mode Classification and Path Prediction with GPS Trajectories

The advancement of positioning technology enables people to use GPS devices to record their location histories. The patterns and contextual information hidden in GPS data opens variety of research issues including trajectory prediction, transportation mode detection, travel route recommendation and many more. Most existing studies have been performed to address these individual issues, but they have the following limitations. 1) Single task learning does not consider the correlations among the correlated tasks. 2) A large number of training samples are required to achieve better performance. In this paper, we propose a semi-supervised multi-task deep learning model to perform the tasks of transportation mode classification and path prediction simultaneously with GPS trajectories. Our model uses both labelled and unlabeled dataset for model training dataset, and concurrently perform both tasks in parallel. Experimental results over a large trajectory dataset collected in Beijing show that our proposal achieves significant performance improvement in terms of all evaluation metrics by comparison with baseline models.