Spatio-Temporal Graph Convolutional Networks: Optimised Temporal Architecture

TL;DR

The paper investigates how to optimise the temporal architecture of spatio-temporal graph convolutional networks (ST-GCNs) for time-series data on graphs. It compares CNN-based and LSTM-based temporal blocks and introduces a novel CNN-GCN-CNN-LSTM hybrid, supported by theoretical arguments and extensive experiments across five datasets. The results show that combining temporal block types yields improvements on several datasets, particularly large ones, and that LSTM blocks can improve robustness to noise and overfitting while CNN blocks speed up training. These findings provide practical guidance for selecting temporal architectures in ST-GCNs and highlight the trade-offs between accuracy, training time, and robustness across diverse graph-time domains.

Abstract

Spatio-Temporal graph convolutional networks were originally introduced with CNNs as temporal blocks for feature extraction. Since then LSTM temporal blocks have been proposed and shown to have promising results. We propose a novel architecture combining both CNN and LSTM temporal blocks and then provide an empirical comparison between our new and the pre-existing models. We provide theoretical arguments for the different temporal blocks and use a multitude of tests across different datasets to assess our hypotheses.

Figures (1)

• Figure 1: An overview of the message passing framework used by Yu et al. in their ST-GCN model Yu_2018. In most input cases $C_i$ = 1, thus the $C_h$ channels are produced from different filters in the CNN. The tensor is then permuted with the channels becoming the vertex features for the GCN with each of the t-k+1 layers passed through in parallel.