Beyond Pairwise Modeling: Towards Efficient and Robust Trajectory Similarity Computation via Representation Learning

ICLR 2026 Conference Submission6459 Authors

Keywords: Trajectory Alignment; Trajectory Representation Learning

TL;DR: Due to the computational redundancy of multi-metric supervision and role-specific repetitive encoding, we propose a novel representation learning framework that transcends pairwise modeling for efficient and robust trajectory similarity computation.

Abstract: Accurate trajectory similarity computation is crucial in ride-sharing applications, where trajectories of varying lengths need to be aligned into a uniform representation. Existing methods suffer from reliance on multi-metric supervision and the role-specific encoding required for triplet loss computation, resulting in inefficient computation. To overcome these issues, we move beyond pairwise modeling and propose a novel representation learning framework to achieve efficient and robust trajectory similarity computation, named Hyper2Edge. Hyper2Edge consists of three main components: (i) Hypergraph-based modeling to represent trajectories as hyperedges, instead of single nodes, preserving sequential and structural details; (ii) Hierarchical trajectory representation learning to capture intra- and inter-trajectory patterns; and (iii) A weighted top-$k$ InfoNCE loss to focus on nearest-neighbor relations, addressing the inefficiencies of triplet loss. Evaluated on two public benchmarks, Hyper2Edge achieves an average absolute gain of 7.42% across all evaluation metrics and an average improvement of 45.9% in accuracy compared to state-of-the-art methods, while maintaining competitive training time per epoch on par with the best-performing methods. The code is available at: https://anonymous.4open.science/r/Hyper2Edge-3D2B.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 6459