Dynamic Localisation of Spatial-Temporal Graph Neural Network

TL;DR

DynAGS introduces a dynamic localised framework for adaptive spatial-temporal graph neural networks, arguing that spatial dependencies evolve over time and data exchange should adapt accordingly. It pivotalizes a lightweight Dynamic Graph Generator with cross-attention to fuse residual historical data with current observations, producing time-varying masks and adaptive graphs that enable personalised localisation. Empirical results across nine real-world datasets and two backbone models demonstrate substantial accuracy gains and dramatic reductions in inter-node communication, validating the time-varying, node-specific approach. The work highlights the practical impact of dynamic spatial modelling for efficient distributed deployment in diverse spatial-temporal applications.

Abstract

Spatial-temporal data, fundamental to many intelligent applications, reveals dependencies indicating causal links between present measurements at specific locations and historical data at the same or other locations. Within this context, adaptive spatial-temporal graph neural networks (ASTGNNs) have emerged as valuable tools for modelling these dependencies, especially through a data-driven approach rather than pre-defined spatial graphs. While this approach offers higher accuracy, it presents increased computational demands. Addressing this challenge, this paper delves into the concept of localisation within ASTGNNs, introducing an innovative perspective that spatial dependencies should be dynamically evolving over time. We introduce \textit{DynAGS}, a localised ASTGNN framework aimed at maximising efficiency and accuracy in distributed deployment. This framework integrates dynamic localisation, time-evolving spatial graphs, and personalised localisation, all orchestrated around the Dynamic Graph Generator, a light-weighted central module leveraging cross attention. The central module can integrate historical information in a node-independent manner to enhance the feature representation of nodes at the current moment. This improved feature representation is then used to generate a dynamic sparse graph without the need for costly data exchanges, and it supports personalised localisation. Performance assessments across two core ASTGNN architectures and nine real-world datasets from various applications reveal that \textit{DynAGS} outshines current benchmarks, underscoring that the dynamic modelling of spatial dependencies can drastically improve model expressibility, flexibility, and system efficiency, especially in distributed settings.

Figures (8)

• Figure 1: An overview of the DynAGS framework with a pre-specified node numbers.
• Figure 2: Test accuracies (MAE) of AGCRN (a) and STG-NCDE (b), localised by DynAGS and AGS, evaluated on transportation datasets (PEMS03,PEMS04,PEMS07,and GLA).The black vertical dashed lines denote the accuracies of AGCRN and STG-NCDE when localised by AGS at a localisation degree of 80%. RMSE and MAPE results are suppressed for the sake of room and can be found in https://github.com/wenyingduan/DynAGS.
• Figure 3: Test accuracies (MAE) of AGCRN (a) and STG-NCDE (b), localised by DynAGS and AGS, evaluated on biosurveillance datasets.The black vertical dashed lines denote the accuracies of AGCRN and STG-NCDE when localised by AGS at a localisation degree of 80%.
• Figure 4: Test accuracies (MAPE) of AGCRN and STG-NCDE, localised by DynAGS and AGS, evaluated on blockchain datasets. The black and yellow vertical dashed lines denote the MAPE of AGCRN and STG-NCDE when localised by AGS at a localisation degree of 80%, respectively. The MAPE of the original models are omitted, as the localised ones are always better.
• Figure 5: Results of personalised localisation.