Source-linked AI summary
GMAN: A Graph Multi-Attention Network for Traffic Prediction
Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, Jianzhong Qi
TL;DR
Long-term traffic prediction is challenging because spatial and temporal correlations are complex and forecast errors can propagate across future steps. GMAN uses attention-based encoder-decoder components and transform attention to model these dependencies, achieving state-of-the-art results on two real-world datasets, including a reported 4% improvement in 1-hour-ahead prediction.
Problem
Long-term traffic prediction lacks satisfactory progress because dynamic spatial correlations, nonlinear temporal correlations, and error propagation make far-future forecasts challenging.
Method
GMAN combines spatial and temporal attention with gated fusion in encoder-decoder blocks, using transform attention to map historical features to future representations.
Results
GMAN achieves state-of-the-art results on two real-world traffic datasets, with a reported 4% improvement over state-of-the-art baselines in 1-hour-ahead prediction.
Takeaways & Limitations
GMAN's advantages become more evident for predictions farther into the future, including the 1-hour-ahead horizon.
Abstract
from arXiv · showhide
Long-term traffic prediction is highly challenging due to the complexity of traffic systems and the constantly changing nature of many impacting factors. In this paper, we focus on the spatio-temporal factors, and propose a graph multi-attention network (GMAN) to predict traffic conditions for time steps ahead at different locations on a road network graph. GMAN adapts an encoder-decoder architecture, where both the encoder and the decoder consist of multiple spatio-temporal attention blocks to model the impact of the spatio-temporal factors on traffic conditions. The encoder encodes the input traffic features and the decoder predicts the output sequence. Between the encoder and the decoder, a transform attention layer is applied to convert the encoded traffic features to generate the sequence representations of future time steps as the input of the decoder. The transform attention mechanism models the direct relationships between historical and future time steps that helps to alleviate the error propagation problem among prediction time steps. Experimental results on two real-world traffic prediction tasks (i.e., traffic volume prediction and traffic speed prediction) demonstrate the superiority of GMAN. In particular, in the 1 hour ahead prediction, GMAN outperforms state-of-the-art methods by up to 4% improvement in MAE measure. The source code is available at https://github.com/zhengchuanpan/GMAN.
Introduction
Long-term traffic prediction is difficult because spatial relationships vary over time, temporal dependencies can be nonlinear, and prediction errors propagate across future steps. GMAN addresses these challenges with attention mechanisms and a transform layer within an encoder-decoder architecture.
- Challenges: Long-term graph-based traffic prediction remains challenging because dynamic spatial correlations and nonlinear temporal correlations vary across time.Nearby sensors may not remain highly correlated, while distant historical time steps can better relate to future conditions.
- Challenges: Small errors at one prediction step can amplify farther into the future, making long-horizon forecasting difficult.
- Approach: GMAN predicts traffic volume and speed on road-network graphs using an encoder-decoder architecture for future time steps.
- Approach: Spatial and temporal attention mechanisms model dynamic spatial and nonlinear temporal correlations, while gated fusion combines their extracted information.
- Approach: A transform attention layer converts encoded historical features into future representations and models direct historical–future relationships to alleviate error propagation.
Related Work
Prior traffic-prediction research combines temporal models with convolutional or graph-based methods, but graph models can generate forecasts step by step and suffer error propagation. Attention-based graph methods offer flexible dependency modeling, while static graph structures remain limited for dynamic spatial relationships.
- Traffic Prediction: Deep learning models such as LSTM generally capture traffic temporal correlations better than traditional time-series and machine-learning methods.
- Traffic Prediction: Graph convolutional networks model non-Euclidean road-network correlations, but step-by-step multi-step prediction can propagate errors across forecast steps.
- Deep Learning on Graphs: Graph deep-learning research extends convolutional networks to arbitrary graphs and also learns low-dimensional vertex representations preserving graph structure.
- Deep Learning on Graphs: A WaveNet-GCN method uses static adjacency matrices, which makes capturing dynamic spatial correlations difficult.
- Attention Mechanism: Attention mechanisms adaptively focus on relevant input features and have been applied to graph-structured data for spatial-correlation modeling.
Preliminaries
The paper represents a road network as a weighted directed graph and traffic observations as graph signals. Given historical signals at all vertices, the forecasting task is to predict the next sequence of traffic conditions across the network.
- Road-Network Representation: A road network is modeled as a weighted directed graph G = (V, E, A), with vertices, connectivity edges, and a weighted adjacency matrix.Vertices represent road-network points such as traffic sensors; adjacency weights represent road-network proximity.
- Traffic Signal: Traffic conditions at time t are represented as a graph signal X_t ∈ R^N×C, where C denotes the number of traffic conditions.Examples include traffic volume and traffic speed.
- Forecasting Problem: Given observations from historical P time steps at N vertices, the task predicts traffic conditions for the next Q time steps at all vertices.
Graph Multi-Attention Network
GMAN models traffic on road-network graphs by combining spatio-temporal embeddings with spatial and temporal attention in an encoder-decoder architecture. Group spatial attention reduces the cost of modeling correlations among many vertices, while transform attention links historical features to future representations.
- Architecture: GMAN uses an encoder-decoder architecture with L ST-Attention blocks in both components, plus spatio-temporal embedding, transform attention, and fully connected layers.The embedding combines spatial and temporal information, while each ST-Attention block uses gated fusion.
- Spatio-Temporal Embedding: Spatio-temporal embedding combines graph-structure information from spatial embeddings with time information from temporal embeddings.Temporal features encode day-of-week and time-of-day before fusion with spatial representations.
- Spatial Attention: Spatial attention dynamically weights vertices at each time step to capture changing correlations between sensors in the road network.The mechanism uses traffic features and graph structure when learning attention scores.
- Spatial Attention: Group spatial attention computes intra-group and inter-group attention to reduce the cost of spatial correlation modeling for large graphs.With group size M = 3√2N, the attention-score count reaches 2^-1/3N^4/3, which is much smaller than N^2.
- Temporal Attention: Temporal attention models non-linear correlations between time steps using traffic features and time context.It restricts attention to earlier time steps to preserve causality.
- Transform Attention: Transform attention directly relates future and historical time steps to convert encoded historical features into future representations for the decoder.The mechanism adaptively selects relevant features across historical time steps before decoding.
Experiments
GMAN is evaluated on two real-world traffic prediction tasks across 15-minute, 30-minute, and 1-hour horizons, using graph-based baselines and standard error metrics. The experiments report stronger long-term performance, fault tolerance, component effectiveness, and efficient multi-step inference.
- Datasets and setup: GMAN is evaluated on traffic volume prediction with 95 Xiamen sensors and traffic speed prediction with 325 PeMS sensors.The datasets use 5-minute time steps, with historical and prediction windows of 12 steps each.
- Forecasting performance: Table 1 compares methods at 15-minute, 30-minute, and 1-hour horizons using MAE, RMSE, and MAPE on the Xiamen and PeMS datasets.The baselines include traditional time-series, machine-learning, recurrent, and graph-based deep-learning methods.
- Forecasting performance: GMAN achieves state-of-the-art performance, with its advantages becoming more evident for long-term predictions such as 1 hour ahead.The paper reports favorable comparisons with Graph WaveNet in long-term traffic prediction.
- Statistical significance: A significance test for 1-hour-ahead prediction yields p-value < 0.01 versus Graph WaveNet, indicating statistically better performance for GMAN.The comparison uses a T-Test.
- Robustness: GMAN is more fault tolerant than state-of-the-art methods when 10% to 90% of historical observations are randomly dropped.The fault-tolerance evaluation makes 1-hour-ahead predictions after replacing randomly selected input values with zeros.
- Ablation study: Removing spatial attention, temporal attention, or gated fusion consistently worsens MAE across prediction steps, supporting the contribution of these components.The ablations are GMAN-NS, GMAN-NT, and GMAN-NG; Figure 9 reports MAE for GMAN and the variants.
- Computation time: GMAN and Graph WaveNet generate 12 prediction steps in one run and require similar training and inference computation costs.STGCN and DCRNN use iterative computation for the 12 prediction results, while DCRNN is slower during training.
Conclusion
The paper proposes GMAN for traffic prediction on road-network graphs using spatial and temporal attention with gated fusion and transform attention. Experiments on two real-world datasets show state-of-the-art results, with larger advantages farther into the prediction horizon.
- Conclusion: GMAN predicts future traffic conditions on road-network graphs using spatial and temporal attention with gated fusion.A transform attention mechanism is designed to ease error propagation and improve long-term prediction.