Spatio‑Temporal Graph Convolution Networks for Traffic Forecasting: Gaode's HSTGCN Approach

The talk, delivered by Gaode senior algorithm expert Ji Chenguang, introduced a novel spatio‑temporal graph convolution algorithm that cleverly exploits massive user navigation planning data to "derive" future congestion, significantly boosting prediction accuracy and showcasing its application within Gaode’s services.

Traffic prediction aims to forecast road conditions at future moments (e.g., 30 minutes ahead) by modeling the spatio‑temporal evolution of congestion. Traditional data‑driven methods often suffer from delayed predictions, especially during sudden congestion, because they rely on coarse event features and cannot provide a minute‑level early‑warning signal.

Gaode’s solution, presented as the Hybrid Spatio‑Temporal Graph Convolution Network (HSTGCN), first estimates future road‑level traffic flow from real‑time ETA updates collected from navigation sessions. These flow estimates are aggregated across all users to obtain a city‑wide future traffic distribution. A domain‑transformer converts flow features into travel‑time features, followed by a 1‑D temporal gated convolution, a graph convolution based on a composite adjacency matrix, additional temporal convolutions, and a fully‑connected decoder that outputs predicted travel times for the entire road network.

The model’s engineering pipeline streams navigation data to a real‑time processing system, merges it with matched trajectory data, and feeds both flow and travel‑time features into HSTGCN for instantaneous forecasting. Compared with the previous state‑of‑the‑art STGCN, HSTGCN better captures sudden congestion, enabling dynamic traffic diversion and reducing violations caused by time‑dependent restrictions.

Beyond prediction, the approach opens a path toward proactive traffic dispatch: future‑aware route planning can feed into traffic‑signal control or toll‑adjustment engines, forming a closed‑loop system that dynamically alleviates congestion, saves travel time, and improves overall transportation efficiency.