How meaningful are similarities in deep trajectory representations?

Finding similar trajectories is an important task in moving object databases. However, classical similarity models face several limitations, including scalability and robustness. Recently, an approach named t2vec proposed transforming trajectories into points in a high dimensional vector space, and this transformation approximately keeps distances between trajectories. t2vec overcomes that scalability limitation: Now it is possible to cluster millions of trajectories. However, the semantics of the learned similarity values – and whether they are meaningful – is an open issue. One can ask: How does the configuration of t2vec affect the similarity values of trajectories? Is the notion of similarity in t2vec similar, different, or even superior to existing models? As for any neural-network-based approach, inspecting the network does not help to answer these questions. So the problem we address in this paper is how to assess the meaningfulness of similarity in deep trajectory representations. Our solution is a methodology based on a set of well-defined, systematic experiments. We compare t2vec to classical models in terms of robustness and their semantics of similarity, using two real-world datasets. We give recommendations which model to use in possible application scenarios and use cases. We conclude that using t2vec in combination with classical models may be the best way to identify similar trajectories. Finally, to foster scientific advancement, we give the public access to all trained t2vec models and experiment scripts. To our knowledge, this is the biggest collection of its kind.