Network-Wide Traffic Volume Estimation from Speed Profiles using a Spatio-Temporal Graph Neural Network with Directed Spatial Attention

AI summary

Overview

• Research area: Machine learning for transportation — spatio-temporal graph neural networks applied to urban traffic volume estimation.
• Technical level: Advanced (assumes familiarity with graph attention, GNNs, and traffic flow theory, though the paper explains its own architecture carefully).
• Scope: The paper introduces HDA-STGNN, a two-branch graph neural network that estimates daily traffic volume profiles on every road segment of a city network using only probe speed data, static road attributes, and road topology — no volume sensors required at inference.

What This Paper Is About

Most traffic volume models either forecast future volumes on roads that already have sensors, or fill in missing volumes by borrowing data from nearby sensors. Both require volume measurements at prediction time, which fails in sensor-scarce cities. This paper instead learns a transformation from widely available data — probe vehicle speed profiles, static road descriptors, and the directed topology of the road network — into daily traffic volume profiles, so that volumes can be estimated on any road segment, including ones never instrumented.

Key Contributions

1. Removal of the volume-data dependency at inference. The model maps daily speed profiles, static attributes, and local topology to daily volume profiles, so no volume observations are needed when making predictions.
2. A hybrid two-branch architecture. The HDA-STGNN processes dynamic speed data through a factorized spatio-temporal branch (1D temporal convolution, Directed Graph Attention, 1D temporal convolution) and static descriptors through a separate spatial branch of DGAT layers, then concatenates the embeddings.
3. Directed spatial attention on the road topology. Directed Graph Attention (DGAT) layers encode edge orientation as incoming, outgoing, or self-loop, with separate projection matrices, attention vectors, and learned directional embeddings for each type, allowing upstream and downstream traffic states to be treated distinctly.
4. A fully inductive design. The learned transformation relies only on local node features and local structure, with no global or graph-specific parameters, making it applicable to unseen road segments and potentially to entirely different road networks without retraining.

Main Findings

• Best overall result: HDA-STGNN achieves RMSE 177.98, MAPE 52.36, mean GEH 6.09, and %GEH>5 of 46.61 on the Lyon Dataset.
• Static attributes matter most: Removing the static descriptor branch (ablation 2) causes the largest degradation, with mean GEH rising to 8.18, RMSE 241.56, MAPE 72.11, and %GEH>5 of 57.94. The authors conclude static features supply structural information about road capacity while speed profiles supply usage information.
• Both branches are needed jointly: Removing the spatio-temporal speed branch (ablation 1) gives RMSE 197.61, MAPE 52.86, GEH 6.53, %GEH>5 49.22.
• Fusion beats a single branch: Merging static and temporal inputs into one branch by duplicating static features across time (ablation 4) yields RMSE 228.15, MAPE 67.74, GEH 7.93, %GEH>5 56.32, worse than the hybrid design.
• Topology helps: Replacing DGAT layers with simple linear transformations (ablation 3) gives RMSE 205.98, MAPE 58.43, GEH 6.78, %GEH>5 52.01.
• Directed attention beats standard GAT on an undirected graph: Ablation 5 gives RMSE 190.44, MAPE 53.21, GEH 6.46, %GEH>5 50.86.
• Intra-day variability remains a gap: Evaluation on raw (non-averaged) volume profiles (ablation 6) gives RMSE 210.76, MAPE 60.88, GEH 6.82, %GEH>5 52.47; the authors suggest richer temporal descriptors such as weather, calendar events, or traffic disruptions could help.
• The 15% GEH criterion is not met: The paper notes a common transportation threshold considers predictions satisfactory when at most 15% of values exceed GEH 5, while the model reports 46.61% exceeding it.
• Known weaknesses: Predictions degrade for road segments whose features diverge from the training distribution, and some segments show significant peak under- or over-estimation, particularly during morning rush hours (8–9 a.m.). Network-wide visualization at 9 a.m. shows plausible spatial patterns (higher volumes on major arteries) but a lack of flow conservation at certain intersections.

Methodology in Plain English

The researchers represent the city's road network as an oriented dual graph: each road segment in each direction becomes a node, and edges represent physical connections between segments, directed according to traffic flow and traffic rules. This avoids the common practice of building a graph over sensor stations, which blurs local flow dynamics.

For each labeled road segment and each day, they build a training sample from the segment's K-hop neighborhood (K = 2), including the local subgraph, the static descriptors of its nodes, and the speed profiles of its nodes for that day. The target is the segment's own volume profile.

The network has two branches. The spatio-temporal branch takes the speed profiles (with time-of-day and weekday encodings added) and applies, for each of K layers, a temporal 1D convolution, then a directed graph attention layer over neighbors, then another temporal 1D convolution — a factorized design that handles space and time separately and is non-causal, since input and output cover the same day. The spatial branch applies K directed graph attention layers to the static descriptors alone. Both branches produce node embeddings that are concatenated, and two fully connected layers turn the combined representation into a predicted volume profile for every node in the subgraph.

Only the target node's predicted profile is compared against ground truth, using a Huber loss with threshold 50, with both profiles rescaled to original units. The model was trained on 522 labeled nodes, with 20% of sensor nodes held out entirely as a validation set, and results averaged over five random seeds.

Data came from 653 Lyon sensors covering all of 2021 with roughly 9% missing values, combined with HERE Maps static descriptors (seven features: speed limit, number of lanes, link length, free flow speed, curvature, slope percent, and functional class) and HERE average speed profiles (96 fifteen-minute intervals per weekday). Because only weekday-averaged speed data was available, volume profiles were averaged by weekday for the main experiment, giving 4,159 unique samples with averaged volumes and 216,885 unique samples with raw volumes.

Why This Matters

The work targets a practical bottleneck in urban traffic analytics: cities with sparse or no fixed sensors still need network-wide volume estimates for planning, management, and environmental assessment. By learning an inductive mapping from broadly available speed and map data, the method offers a path to full network coverage where imputation and forecasting methods cannot operate.

Real-world applications:

• Air quality and emissions assessment, which the paper cites as a motivation requiring network-wide traffic volumes.
• Urban planning and traffic management on roads that have no instrumentation.
• Rapid deployment in sensor-scarce cities where probe speed and map data are available but volume counts are not.
• Transfer to entirely different road networks without architectural changes or retraining, thanks to the inductive design.

Industry relevance: Geospatial data providers, mapping platforms, and mobility analytics vendors already distribute probe speed and static road attributes at scale, so a model that converts these into volume estimates could extend commercial traffic-data products into cities where installing and maintaining sensor infrastructure is uneconomical.

Future Directions

• Increase the variability and diversity of the training dataset to improve out-of-distribution generalization, since some road segments fail when their feature profiles differ from the training distribution.
• Incorporate physical constraints such as flow conservation to improve local coherence at intersections, where the network-wide visualization shows inconsistencies.
• Add richer temporal descriptors such as weather, calendar events, or traffic disruptions to capture intra-day volume variability, which the raw-volume ablation (6) suggests is currently under-modeled.
• Test transfer to unseen road networks and sensor-free areas, which the inductive design theoretically permits but which the paper does not evaluate.

Target Audience

Researchers and graduate students working on spatio-temporal graph neural networks, traffic state estimation, or spatio-temporal imputation; transportation engineers and urban mobility analysts interested in network-wide volume estimation without sensor coverage; and practitioners at geospatial or mobility data companies evaluating whether probe speed data can substitute for loop detectors. Readers need some background in graph neural networks and attention mechanisms, though the paper explains its architectural choices in detail.

Authors’ abstract

Existing traffic volume estimation methods typically address either forecasting traffic on sensor-equipped roads or spatially imputing missing volumes using nearby sensors. While forecasting models generally disregard unmonitored roads by design, spatial imputation methods explicitly address network-wide estimation; yet this approach relies on volume data at inference time, limiting its applicability in sensor-scarce cities. Unlike traffic volume data, probe vehicle speeds and static road attributes are more broadly accessible and support full coverage of road segments in most urban networks. In this work, we present the Hybrid Directed-Attention Spatio-Temporal Graph Neural Network (HDA-STGNN), an inductive deep learning framework designed to tackle the network-wide volume estimation problem. Our approach leverages speed profiles, static road attributes, and road network topology to predict daily traffic volume profiles across all road segments in the network. To evaluate the effectiveness of our approach, we perform extensive ablation studies that demonstrate the model's capacity to capture complex spatio-temporal dependencies and highlight the value of topological information for accurate network-wide traffic volume estimation without relying on volume data at inference time.