Abstract: Quantifying the similarity between two trajectories is a fundamental research that underlies many trajectory-based applications. Conventional methods suffer from inefficiency and noise sensitivity, making it difficult to achieve large-scale deployments. Current researches start to explore the emerging deep representation learning method, which maps high-dimensional trajectory data to a low-dimensional vector space for efficiently performing similarity measurement by computing the distance between trajectory representations. This paper pioneers the idea of Transformer, and proposes a geography-aware deep representation learning model for trajectory similarity computation: First, the two-dimensional coordinate point is converted into a one-dimensional sequence using Geohash algorithm, which can preserve the spatial correlations of the trajectory point during the embedding. Second, a deep trajectory representation learning model is constructed based on the Transformer framework, and a masked point strategy is employed to ensure that the model can acquire robust vector representations from low-frequency, noisy data. Final, a geography-aware loss function is devised to penalize the model and narrow the representation of spatially similar trajectories via a distance factor. Experiments show that the proposed method outperforms the state-of-the-art model in the similarity measurement and is at least one order of magnitude faster than the traditional models in terms of computational efficiency.

Key words: deep representation learning, trajectory representation, Transformer model, similarity measurement