Hybrid Spatio‑Temporal Graph Convolutional Network for Precise Traffic Prediction

The 2020 Yunqi Conference (Sept 17‑18) featured a "Smart Mobility" session organized by Alibaba Amap, where senior algorithm expert Ji Chenguang presented the paper "Hybrid Spatio‑Temporal Graph Convolutional Network: More Accurate Traffic Prediction".

Ji explained how Amap leverages massive real‑time navigation data to predict future road congestion, a core technology behind features such as congestion‑avoidance routing and ETA estimation for its 530 million‑plus monthly active users.

The talk was divided into three parts: (1) What traffic prediction is; (2) The Hybrid Spatio‑Temporal Graph Convolutional Network (HSTGCN) that can "derive" future traffic conditions; (3) Future applications, from prediction to traffic scheduling.

What traffic prediction is

Traffic prediction aims to forecast the traffic state (e.g., congestion level) at future time points based on current observations. Two main approaches exist: (a) simulation‑based methods that combine vehicle origin‑destination data with traffic dynamics theory, and (b) data‑driven methods that learn statistical relationships from historical traffic data. Existing data‑driven models (random forest, GBDT, early deep learning models, and recent convolutional or graph‑convolutional networks) often suffer from delayed predictions because they lack fine‑grained, minute‑level traffic flow signals.

Ji introduced Amap’s solution, which was published at KDD 2020.

Hybrid Spatio‑Temporal Graph Convolutional Network (HSTGCN)

Step 1: From massive real‑time navigation sessions, Amap estimates each vehicle’s future contribution to road traffic flow using the ETA feedback loop. Aggregating these contributions per road yields a forecast of future traffic volume across the entire network.

Step 2: A novel domain‑converter transforms the predicted traffic‑flow features into road travel‑time features. A 1‑D temporal gated convolution extracts higher‑level temporal patterns, followed by a graph convolution network built on a composite adjacency matrix that captures spatial dependencies among road segments. Two additional temporal gated convolutions and a fully‑connected decoder produce the final prediction of future travel times for the whole road network.

HSTGCN effectively learns the traffic dynamics linking increased flow to congestion, combining the strengths of simulation‑based (knowledge‑driven) and data‑driven (statistical) methods.

The deployment architecture streams navigation data in real time, processes it with Blink for flow estimation, matches GPS traces to the road network for travel‑time extraction, and feeds both feature streams into HSTGCN for instantaneous traffic‑state prediction. The predictions are then used to refine ETA calculations and route planning.

Experimental results show that HSTGCN outperforms the previous state‑of‑the‑art STGCN, especially in predicting sudden congestion spikes.

Application Outlook: From Prediction to Traffic Scheduling

Unlike passive statistical forecasts, HSTGCN enables proactive traffic management. Predicted future congestion can feed into dynamic traffic‑control modules such as adaptive traffic‑signal timing or highway toll adjustments, which in turn influence subsequent route planning, forming a closed‑loop intelligent traffic system.

This real‑time, dynamic operation promises to alleviate congestion, reduce travel time, and improve overall transportation efficiency.