GMAN: A Graph Multi-Attention Network for Traffic Prediction

Zheng, Chuanpan; Fan, Xiaoliang; Wang, Cheng; Qi, Jianzhong · 2020 · OpenAlex-citations

DOI: 10.1609/aaai.v34i01.5477

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of long-term traffic prediction, specifically focusing on modeling complex spatio-temporal correlations and mitigating error propagation in road networks. While existing methods using Graph Convolutional Networks (GCNs) and Recurrent Neural Networks (RNNs) perform well for short-term predictions, they struggle with long-term horizons due to dynamic spatial dependencies, non-linear temporal fluctuations, and the accumulation of errors across sequential prediction steps. To overcome these limitations, the authors propose GMAN (Graph Multi-Attention Network), a model designed to predict traffic conditions, such as volume and speed, multiple time steps ahead. GMAN utilizes an encoder-decoder architecture composed of stacked Spatio-Temporal Attention (ST-Attention) blocks. The model incorporates a Spatio-Temporal Embedding (STE) module that fuses static graph structure information, derived via node2vec, with dynamic temporal features encoded through one-hot representations of time-of-day and day-of-week. Within each ST-Attention block, a spatial attention mechanism dynamically assigns weights to neighboring sensors to capture time-variant spatial correlations, while a temporal attention mechanism models non-linear dependencies between different historical time steps. These representations are adaptively fused using a gated mechanism. Crucially, GMAN introduces a Transform Attention layer between the encoder and decoder. This layer establishes direct relationships between historical inputs and future prediction steps, allowing the decoder to generate future representations directly from historical data rather than relying solely on previous predictions, thereby alleviating error propagation. To handle computational complexity in large graphs, the authors also propose a group spatial attention mechanism that partitions vertices to reduce attention score calculations. The model was evaluated on two real-world datasets: a local dataset from Xiamen, China, with 95 sensors, and the PeMS dataset with 325 sensors. The experiments focused on traffic volume and speed prediction tasks. The results demonstrate that GMAN outperforms state-of-the-art baseline methods, including those based on GCNs and WaveNet. Specifically, for one-hour ahead predictions, GMAN achieved up to a 4% improvement in Mean Absolute Error (MAE) compared to existing approaches. The study highlights that the attention-based mechanisms effectively capture the dynamic and non-linear nature of traffic data, providing superior accuracy and fault tolerance in long-term forecasting scenarios.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation