Spatio-temporal meta-learning for trajectory representation learning

doi:10.1016/j.knosys.2025.114141

Spatio-temporal meta-learning for trajectory representation learning

2025-07-23

0

PRE

AI

DOI：10.1016/j.knosys.2025.114141

Original

Original request for help

Save

Abstract

En 中文

Trajectory representation learning translates sequences into low-dimensional vectors that are convenient for computer processing and analysis. Trajectory representation learning is widely used by various intelligent applications and is notable for its ability to enhance application performance. However, previous methods assume that trajectories are independently and identically distributed in time and space. In practice, trajectories exhibit significant heterogeneity in time and space sources due to the uncertainty of individual activities and the diversity of activity patterns. This leads to bias in the generated representation vectors that fail to effectively support various geographic applications. To address this issue, this study proposes a spatio-temporal meta-learning method for trajectory representation learning, namely STMetaT, which aims to generate accurate trajectory representation vectors. STMetaT designs a spatio-temporal constraint sampling module that divides trajectory sets into subsets based on the frequency and density of trajectories, which constructs training task samples with diverse spatio-temporal semantics. And, STMetaT uses a multi-view local encoder to generate representation vectors for each subset by fusing the diversity of trajectory semantics. Finally, STMetaT learns a generalization process from local to global promotability to optimize the trajectory representation vectors. Extensive experiments on two urban trajectory datasets show that STMetaT outperforms baseline methods in three classical evaluation tasks, thereby improving the performance of trajectory representation. The proposed method provides an approach for learning trajectory representation by combining meta-learning, and also provides a methodological reference for various intelligent applications.

Keywords:

trajectory representation learning

spatio-temporal meta-learning

heterogeneous trajectories

multi-view local encoder

geographic applications

Spatio-temporal meta-learning for trajectory representation learning

Abstract

Journal

Organization