Hybrid Spatio-Temporal Graph Convolutional Network (H‑STGCN) for Traffic Forecasting

Overview – Spatio‑temporal prediction is crucial for applications such as weather forecasting and transportation planning. Traffic forecasting, a typical spatio‑temporal problem, is challenging because traditional models rely only on travel‑time features and cannot capture the overall traffic condition. The authors (Gaode Machine Learning team) propose a Hybrid Spatio‑Temporal Graph Convolutional Network (H‑STGCN) that incorporates planned traffic flow derived from navigation data, significantly improving prediction performance (paper accepted at KDD 2020).

Planned Traffic Flow – The authors extract "planned traffic flow" from Gaode’s navigation engine, which reflects users' travel intentions and future traffic volume. This signal is richer than conventional travel‑time features and provides finer granularity than event‑level attributes.

Domain Transformer – To integrate the heterogeneous flow signal into a travel‑time prediction model, a novel domain transformer is designed. It converts flow information into travel‑time signals using two stacked layers: a segment‑wise convolution (capturing road‑specific details) and a shared convolution (capturing global triangular flow‑time mapping).

Graph Convolution with Compound Adjacency Matrix – Because road networks are non‑Euclidean, the model employs graph convolutions to capture spatial dependencies. A compound adjacency matrix is introduced, extending the traditional distance‑decay matrix with covariance of travel times between road segments, better reflecting true traffic proximity.

H‑STGCN Architecture – The overall framework consists of two input tensors (ideal future flow and travel‑time), a domain transformer (module a), two independent gated temporal convolutions (module b), a graph convolution based on the compound adjacency matrix (module c), additional gated convolutions, and a final fully‑connected layer for prediction.

Data Processing – Ideal future flow is approximated from online navigation data. The navigation engine provides per‑segment planned flow (volf) for future time steps. Historical travel‑time tensors and their corresponding historical averages are also fed to the model.

Training Details – Data augmentation adds Gaussian noise to flow channels. The loss function is L1 (MAE) between predicted and ground‑truth travel times.

Experiments – Two real‑world datasets (W3‑715 and E5‑2907) covering thousands of road segments are used. Baselines include Historical Average, Linear Regression, GBRT, MLP, Seq2Seq, and STGCN. Variants such as STGCN with the compound adjacency matrix (STGCN (Im)) and H‑STGCN with flow set to 1 (H‑STGCN (1)) are also evaluated. H‑STGCN consistently outperforms all baselines on MAE, MAPE, and RMSE, especially in sudden‑congestion scenarios.

Case Study – A sudden‑congestion example on a highway shows that H‑STGCN can anticipate congestion up to 30 minutes in advance, whereas models without future flow lag behind.

Scalability – Inference time per sub‑network (a few thousand road segments) is under 100 ms, enabling online deployment by partitioning city road networks.

Impact – H‑STGCN has been deployed in Gaode’s ETA prediction, reducing severe‑error cases by 15 % and offering a data‑driven way to model the interaction between user intent and traffic evolution. It holds promise for proactive traffic management such as intelligent traffic‑light control and dynamic road‑pricing.

References – The summary includes key citations to related works on diffusion convolutional recurrent networks, traffic flow prediction with big data, STGCN, and reinforcement‑learning traffic‑light control.