TremBR: Exploring road networks for trajectory representation learning

In this article, we propose a novel representation learning framework, namely TRajectory EMBedding via Road networks (Trembr), to learn trajectory embeddings (low-dimensional feature vectors) for use in a variety of trajectory applications. The novelty of Trembr lies in (1) the design of a recurrent neural network-(RNN) based encoder-decoder model, namely Traj2Vec, that encodes spatial and temporal properties inherent in trajectories into trajectory embeddings by exploiting the underlying road networks to constrain the learning process in accordance with the matched road segments obtained using road network matching techniques (e.g., Barefoot [24, 27]), and (2) the design of a neural network-based model, namely Road2Vec, to learn road segment embeddings in road networks that captures various relationships amongst road segments in preparation for trajectory representation learning. In addition to model design, several unique technical issues raising in Trembr, including data preparation in Road2Vec, the road segment relevance-aware loss, and the network topology constraint in Traj2Vec, are examined. To validate our ideas, we learn trajectory embeddings using multiple large-scale real-world trajectory datasets and use them in three tasks, including trajectory similarity measure, travel time prediction, and destination prediction. Empirical results show that Trembr soundly outperforms the state-of-the-art trajectory representation learning models, trajectory2vec and t2vec, by at least one order of magnitude in terms of mean rank in trajectory similarity measure, 23.3% to 41.7% in terms of mean absolute error (MAE) in travel time prediction, and 39.6% to 52.4% in terms of MAE in destination prediction.