Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

T-JEPA: A Joint-Embedding Predictive Architecture for Trajectory Similarity Computation

Lihuan Li, Hao Xue, Yang Song, Flora Salim

TL;DR

The paper tackles robust trajectory similarity learning from unlabeled GPS data. It introduces T-JEPA, a JEPA-based framework that automates augmentation through representation-space resampling and employs a predictive module to learn high-level trajectory semantics. An AdjFuse module enriches contextual information by aggregating adjacent regional features via a sliding kernel over grid-cell embeddings. Empirical results on five urban GPS and FourSquare datasets demonstrate competitive or superior performance to state-of-the-art baselines and show strong robustness to irregular sampling, illustrating a scalable self-supervised approach for trajectory similarity with practical impact on routing, clustering, and pattern discovery.

Abstract

Trajectory similarity computation is an essential technique for analyzing moving patterns of spatial data across various applications such as traffic management, wildlife tracking, and location-based services. Modern methods often apply deep learning techniques to approximate heuristic metrics but struggle to learn more robust and generalized representations from the vast amounts of unlabeled trajectory data. Recent approaches focus on self-supervised learning methods such as contrastive learning, which have made significant advancements in trajectory representation learning. However, contrastive learning-based methods heavily depend on manually pre-defined data augmentation schemes, limiting the diversity of generated trajectories and resulting in learning from such variations in 2D Euclidean space, which prevents capturing high-level semantic variations. To address these limitations, we propose T-JEPA, a self-supervised trajectory similarity computation method employing Joint-Embedding Predictive Architecture (JEPA) to enhance trajectory representation learning. T-JEPA samples and predicts trajectory information in representation space, enabling the model to infer the missing components of trajectories at high-level semantics without relying on domain knowledge or manual effort. Extensive experiments conducted on three urban trajectory datasets and two Foursquare datasets demonstrate the effectiveness of T-JEPA in trajectory similarity computation.

T-JEPA: A Joint-Embedding Predictive Architecture for Trajectory Similarity Computation

TL;DR

The paper tackles robust trajectory similarity learning from unlabeled GPS data. It introduces T-JEPA, a JEPA-based framework that automates augmentation through representation-space resampling and employs a predictive module to learn high-level trajectory semantics. An AdjFuse module enriches contextual information by aggregating adjacent regional features via a sliding kernel over grid-cell embeddings. Empirical results on five urban GPS and FourSquare datasets demonstrate competitive or superior performance to state-of-the-art baselines and show strong robustness to irregular sampling, illustrating a scalable self-supervised approach for trajectory similarity with practical impact on routing, clustering, and pattern discovery.

Abstract

Trajectory similarity computation is an essential technique for analyzing moving patterns of spatial data across various applications such as traffic management, wildlife tracking, and location-based services. Modern methods often apply deep learning techniques to approximate heuristic metrics but struggle to learn more robust and generalized representations from the vast amounts of unlabeled trajectory data. Recent approaches focus on self-supervised learning methods such as contrastive learning, which have made significant advancements in trajectory representation learning. However, contrastive learning-based methods heavily depend on manually pre-defined data augmentation schemes, limiting the diversity of generated trajectories and resulting in learning from such variations in 2D Euclidean space, which prevents capturing high-level semantic variations. To address these limitations, we propose T-JEPA, a self-supervised trajectory similarity computation method employing Joint-Embedding Predictive Architecture (JEPA) to enhance trajectory representation learning. T-JEPA samples and predicts trajectory information in representation space, enabling the model to infer the missing components of trajectories at high-level semantics without relying on domain knowledge or manual effort. Extensive experiments conducted on three urban trajectory datasets and two Foursquare datasets demonstrate the effectiveness of T-JEPA in trajectory similarity computation.
Paper Structure (19 sections, 4 equations, 8 figures, 5 tables)

This paper contains 19 sections, 4 equations, 8 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Trajectory similarity computation
  4. Joint-Embedding Predictive Architecture
  5. Preliminaries
  6. Method
  7. Model Overview
  8. Main Components of T-JEPA
  9. Contextual Feature Enrichment
  10. Experiments
  11. Experimental Settings
  12. Quantitative Evaluation
  13. Most similar trajectory search
  14. Robustness
  15. Approximating heuristic measures
  16. ...and 4 more sections

Figures (8)

  • Figure 1: Given a trajectory at the bottom, we illustrate the structural differences between the contrastive learning framework (pink arrow branch) and the JEPA framework (orange arrow branch). On the top, we show the 5-NN results of TrajCL chang2023contrastive (left figure) using contrastive learning and our proposed T-JEPA (right figure) after fine-tuning by Hausdorff measures. The query trajectory is in red and the matched trajectories are orange heatmaps.
  • Figure 2: The structure of JEPA framework.
  • Figure 3: Our proposed T-JEPA framework. Given a trajectory, this training strategy is designed to predict multiple sampled target embeddings from one sampled context trajectory.
  • Figure 4: The illustration of the AdjFuse module.
  • Figure 5: The study areas (in light blue) of the datasets used in our paper.
  • ...and 3 more figures