HLGST: Hybrid local–global spatio-temporal model for travel time estimation using Siamese graph convolutional with triplet networks

Abstract

Travel time estimation (TTE) is a crucial and challenging task due to the complex spatial and dynamic temporal correlations between local and global traffic regions. Though many existing methods have used multi-graph structures to model traffic variations, the majority of them are unable to capture such sophisticated local and global perspectives dynamically at different time-intervals. Furthermore, these methods have used shallow graphs (i.e., based on speed or distance between traffic nodes), thereby limiting their ability to learn the latent dependencies. To overcome these limitations, we propose a novel Hybrid Local-Global Spatio-Temporal (HLGST) framework for TTE. Specifically, we first introduce a dynamic composition unit that builds local traffic information and multi-dynamic semantic graphs based on the similarities between nodes. Then, the HLGST model learns the local and global dependencies via hybrid correlation method. It mainly comprises two modules: (1) local correlation module that integrates casual TCN layers with a self-attention mechanism to capture local dependencies of traffic patterns. (2) triplet-siamese GCN (TS-GCN) module that is employed to smoothly capture global dependencies based on triplet relationships between multi-dynamical semantic graphs. Moreover, we build a dynamic adaptive learning algorithm to transfer the gained knowledge and leverage it to model multi-objective prediction tasks collaboratively. Extensive experiments conducted on two real-world traffic datasets (Xi’an and Chengdu) demonstrate that our HLGST model outperforms compared baselines and achieves significant improvement.