Transferable multi-level spatial-temporal graph neural network for adaptive multi-agent trajectory prediction

Yu Sun

doi:10.1016/j.knosys.2026.115451

Transferable multi-level spatial-temporal graph neural network for adaptive multi-agent trajectory prediction

2026-01-31

0

PRE

AI

DOI：10.1016/j.knosys.2026.115451

Original

Original request for help

Save

Abstract

En 中文

• A novel transferable multi-level spatiotemporal graph neural network (T-MLSTG) framework is proposed to address the challenges of data distribution shift in real traffic scenarios for multi-agent trajectory prediction. • Windowed mean gradient discrepancy (WMGD) is a novel metric proposed to measure distribution shifts, which explicitly captures the temporal dynamics and phased evolution patterns through gradient statistics within sliding windows. This approach provides a more sensitive and stable criterion for spatiotemporal feature alignment. • We designed a multi-level spatial-temporal graph neural network to learn motion and interaction representations. Its structured three-level design enables comprehensive modeling from individual agents to multi-agent interactions, achieving both accurate prediction within a single domain and effective extraction of domain-invariant features for robust cross-domain adaptation.

Keywords:

multi-agent trajectory prediction

spatiotemporal graph neural network

domain adaptation

gradient discrepancy

motion representation

Transferable multi-level spatial-temporal graph neural network for adaptive multi-agent trajectory prediction

Abstract

Journal

Organization