Temporal Graph Convolutional Network

A hybrid deep learning framework that integrates Graph Convolutional Networks and Gated Recurrent Units to map complex spatial-temporal dependencies.

T-GCN addresses the limitations of traditional time-series models by capturing both network topology and dynamic patterns simultaneously. First introduced by Zhao et al. (2019), the architecture utilizes a Graph Convolutional Network (GCN) to extract spatial features from a graph's Laplacian matrix and a Gated Recurrent Unit (GRU) to track temporal evolution. This dual approach is highly effective for urban infrastructure: benchmarks on the SZ-taxi and Los-loop datasets demonstrate that T-GCN consistently outperforms baseline models like ARIMA and Support Vector Regression. By processing historical traffic speeds or power grid fluctuations as nodal attributes, the system provides accurate long-term forecasting for interconnected high-dimensional data.

https://arxiv.org/abs/1811.05320