[Paper Review] ST-UNet: A Spatio-Temporal U-Network for Graph-structured Time Series Modeling

Bing Yu, Haoteng Yin+1 more|arXiv (Cornell University)|Mar 13, 2019

Advanced Graph Neural Networks26 references50 citations

TL;DR

ST-UNet introduces a Spatio-Temporal U-Net with ST-Pool/ST-Unpool and GCGRU to model and forecast graph-structured time series across multiple spatial and temporal scales, achieving state-of-the-art results on traffic datasets.

ABSTRACT

The spatio-temporal graph learning is becoming an increasingly important object of graph study. Many application domains involve highly dynamic graphs where temporal information is crucial, e.g. traffic networks and financial transaction graphs. Despite the constant progress made on learning structured data, there is still a lack of effective means to extract dynamic complex features from spatio-temporal structures. Particularly, conventional models such as convolutional networks or recurrent neural networks are incapable of revealing the temporal patterns in short or long terms and exploring the spatial properties in local or global scope from spatio-temporal graphs simultaneously. To tackle this problem, we design a novel multi-scale architecture, Spatio-Temporal U-Net (ST-UNet), for graph-structured time series modeling. In this U-shaped network, a paired sampling operation is proposed in spacetime domain accordingly: the pooling (ST-Pool) coarsens the input graph in spatial from its deterministic partition while abstracts multi-resolution temporal dependencies through dilated recurrent skip connections; based on previous settings in the downsampling, the unpooling (ST-Unpool) restores the original structure of spatio-temporal graphs and resumes regular intervals within graph sequences. Experiments on spatio-temporal prediction tasks demonstrate that our model effectively captures comprehensive features in multiple scales and achieves substantial improvements over mainstream methods on several real-world datasets.

Motivation & Objective

• Motivate and model dynamic spatio-temporal graphs where both spatial and temporal dependencies vary over time.
• Develop a multi-scale U-shaped architecture to capture local and global spatio-temporal patterns.
• Propose efficient spatio-temporal pooling and unpooling operators on graphs.
• Leverage graph convolutional gated recurrent units as the backbone for sequence modeling.
• Demonstrate improved prediction accuracy on real-world traffic datasets and validate component effectiveness.

Proposed method

• Generalize U-Net to spatio-temporal graphs with a Spatio-Temporal U-Net (ST-UNet).
• Use graph convolutional gated recurrent units (GCGRU) as the backbone to model temporal dynamics on graphs.
• Introduce ST-Pool to coarsen graphs via deterministic partitioning (gPartition) and aggregate temporal features with dilated recurrent skip connections.
• Introduce ST-Unpool to restore the original graph structure and temporal dependencies, with three upsampling strategies (direct copy, ordered deconv, weighted deconv).
• Fuse high-level pooled features with upsampled outputs through skip connections for multi-scale feature fusion.
• Employ multi-scale feature learning to predict future node attributes or entire graphs on short-horizon forecasts.

Experimental results

Research questions

• RQ1 Can a multi-scale, U-shaped architecture be effectively applied to graph-structured time series to capture both local and global spatio-temporal patterns?
• RQ2 Do spatio-temporal pooling and unpooling operations improve predictive accuracy over flat or single-scale models?
• RQ3 How do dilated recurrent skip-connections and GCGRU backbones contribute to multi-resolution temporal modeling on dynamic graphs?
• RQ4 Which upsampling strategy in ST-Unpool yields the best balance of accuracy and robustness?
• RQ5 What is ST-UNet’s performance and scalability on large-scale graph-structured time series tasks like traffic networks?

Key findings

• ST-UNet consistently outperforms baselines (GCGRU, STGCN, DCRNN) on spatio-temporal traffic prediction across METR-LA and PeMS datasets.
• ST-UNet achieves best MAE, MAPE, and RMSE across 15, 30, and 60 minute horizons on both datasets.
• Ablation studies show ST-Pool and ST-Unpool contribute to performance gains, with full ST-UNet yielding the best results.
• Among upsampling strategies, Direct Copy generally performs best, especially for longer horizons.
• ST-UNet demonstrates scalability benefits on large-scale graphs (PeMS-L), where traditional GCN-based models struggle.