Abstract:Dynamic spatial graph construction is a challenge in graph neural network (GNN) for time series data problems. Although some adaptive graphs are conceivable, only a 2D graph is embedded in the network to reflect the current spatial relation, regardless of all the previous situations. In this work, we generate a spatial tensor graph (STG) to collect all the dynamic spatial relations, as well as a temporal tensor graph (TTG) to find the latent pattern along time at each node. These two tensor graphs share the same nodes and edges, which leading us to explore their entangled correlations by Projected Entangled Pair States (PEPS) to optimize the two graphs. We experimentally compare the accuracy and time costing with the state-of-the-art GNN based methods on the public traffic datasets.

What problem does this paper attempt to address?

The main problems that this paper attempts to solve are the construction of dynamic spatial graphs in time - series data and the mining of the potential time - relatedness of each node, especially in the application of traffic prediction tasks. Specifically, the paper focuses on two main challenges: 1. **Construction of Dynamic Spatial Graphs**: - Existing graph neural network (GNN) methods usually use static spatial graphs and one - dimensional convolutions to process time - series data, which cannot reflect the changes of nodes and edges over time. - To solve this problem, the paper proposes to construct a **Spatial Tensor Graph (STG)**, collecting all dynamic spatial relationships together and extracting more refined features through high - dimensional convolutions. 2. **Mining the Potential Time - relatedness of Each Node**: - Currently popular GNN methods mainly use the spatial structure of the data set to construct graphs, ignoring the time - relatedness of each node. - The paper captures the potential patterns between different time steps of each node by constructing a **Temporal Tensor Graph (TTG)** to improve the performance that only depends on the spatial graph. To optimize these two graphs simultaneously, the paper introduces **Projected Entangled Pair States (PEPS)**, which is a tensor network method that can reduce the number of parameters and improve computational efficiency. Finally, the paper proposes a method named **DSTGNN (Dynamic Spatiotemporal GNN with Tensor Network)** for traffic prediction tasks and conducts experimental verification on multiple public data sets. ### Formula Summary - **Weight Calculation of Spatial Tensor Graph (STG)**: \[ A[i,j,t] = \begin{cases} \exp\left(-\frac{d_{ijt}^2}{\sigma^2}\right), & \text{if } i,j \text{ and } \exp\left(-\frac{d_{ijt}^2}{\sigma^2}\right) \geq \epsilon \\ 0, & \text{otherwise} \end{cases} \] where \( d_{ijt} \) is the sample difference between stations \( i \) and \( j \) at time \( t \), and \( \sigma^2 \) and \( \epsilon \) are set to 0.1 and 0.5 respectively. - **Output Calculation of Spatiotemporal Graph Convolutional Layer (STGCL)**: \[ X_{\text{out}} = B X W_B + A X W_A + B X W_B \] where \( W_A \in \mathbb{R}^{C_i \times K_A \times C_o \times T} \), \( W_B \in \mathbb{R}^{C_i' \times K_B \times C_o' \times N} \), and \( C_i, C_i' \) are input dimensions, \( C_o, C_o' \) are output dimensions. Through these methods, the paper effectively solves the problems of dynamic spatial graph construction and time - relatedness mining and shows its superior performance in the experiments.

Dynamic Spatiotemporal Graph Neural Network with Tensor Network

STGSN — A Spatial–Temporal Graph Neural Network framework for time-evolving social networks

Spatial-temporal Graph Convolutional Networks with Diversified Transformation for Dynamic Graph Representation Learning

Dynamic Spatio-Temporal Graph Network with Adaptive Propagation Mechanism for Multivariate Time Series Forecasting

Adaptive Spatio-temporal Graph Neural Network for traffic forecasting

Spatio-Temporal Graph-TCN Neural Network for Traffic Flow Prediction

Temporal Graph Transformer for Dynamic Network

A dynamical spatial-temporal graph neural network for traffic demand prediction

Spatial-temporal graph neural network based on gated convolution and topological attention for traffic flow prediction

Dynamic Personalized Graph Neural Network with Linear Complexity for Multivariate Time Series Forecasting

Dynamic spatial aware graph transformer for spatiotemporal traffic flow forecasting

Spatial-Temporal Interactive Graph Neural Network for Traffic Forecasting.

Temporal Graph Networks for Deep Learning on Dynamic Graphs

GraphTTE: Travel Time Estimation Based on Attention-Spatiotemporal Graphs.

STG-Mamba: Spatial-Temporal Graph Learning via Selective State Space Model

Dynamic Graph Learning-Neural Network for Multivariate Time Series Modeling

Adaptive Sampling Temporal Graph Network

Analysis of different temporal graph neural network configurations on dynamic graphs

Short-term Load Forecasting Using Spatial-Temporal Embedding Graph Neural Network

Spatio-Temporal Pivotal Graph Neural Networks for Traffic Flow Forecasting

Tensor Graph Convolutional Network for Dynamic Graph Representation Learning