[Paper Review] Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting

Lei Bai, Lina Yao+3 more|arXiv (Cornell University)|Jul 6, 2020

Traffic Prediction and Management Techniques46 references697 citations

TL;DR

The paper introduces AGCRN, an adaptive framework that learns node-specific GCN parameters and data-driven graphs to forecast multi-step traffic without predefined spatial graphs, improving accuracy over state-of-the-art methods.

ABSTRACT

Modeling complex spatial and temporal correlations in the correlated time series data is indispensable for understanding the traffic dynamics and predicting the future status of an evolving traffic system. Recent works focus on designing complicated graph neural network architectures to capture shared patterns with the help of pre-defined graphs. In this paper, we argue that learning node-specific patterns is essential for traffic forecasting while the pre-defined graph is avoidable. To this end, we propose two adaptive modules for enhancing Graph Convolutional Network (GCN) with new capabilities: 1) a Node Adaptive Parameter Learning (NAPL) module to capture node-specific patterns; 2) a Data Adaptive Graph Generation (DAGG) module to infer the inter-dependencies among different traffic series automatically. We further propose an Adaptive Graph Convolutional Recurrent Network (AGCRN) to capture fine-grained spatial and temporal correlations in traffic series automatically based on the two modules and recurrent networks. Our experiments on two real-world traffic datasets show AGCRN outperforms state-of-the-art by a significant margin without pre-defined graphs about spatial connections.

Motivation & Objective

• Motivate learning node-specific patterns beyond shared parameters in GCNs for traffic data.
• Eliminate dependence on hand-crafted adjacency graphs by inferring relationships from data.
• Integrate node-adaptive GCN with recurrent networks to capture spatial-temporal dynamics.
• Provide end-to-end trainable modules that improve multi-step traffic prediction performance.

Proposed method

• Propose Node Adaptive Parameter Learning (NAPL) to generate node-specific GCN parameters from a small node embedding and a shared weight pool.
• Propose Data Adaptive Graph Generation (DAGG) to infer inter-node dependencies via learned node embeddings andSoftmax-ReLU normalization, producing a data-driven adjacency representation.
• Integrate NAPL and DAGG with a GRU-based recurrent backbone to form AGCRN, sharing a unified node embedding across layers.
• Stack AGCRN layers for multi-step forecasting and project outputs with a linear layer.
• Train end-to-end using L1 loss over the next τ steps with Adam optimizer.

Experimental results

Research questions

• RQ1 Can node-specific parameterization improve GCN-based traffic forecasting compared to shared-parameter GCNs?
• RQ2 Is it possible to infer spatial dependencies directly from data without a predefined graph while maintaining predictive accuracy?
• RQ3 Does unifying node embeddings across adaptive modules improve performance and interpretability?
• RQ4 How does AGCRN perform for multi-step traffic forecasting on real-world datasets compared to existing baselines?
• RQ5 What is the impact of embedding dimension on performance and model complexity?

Key findings

• AGCRN outperforms state-of-the-art baselines on PeMSD4 and PeMSD8 in MAE, RMSE, and MAPE across 12 horizons, with notable gains (e.g., MAE reduction to 19.83 on PeMSD4 and 15.95 on PeMSD8).
• Ablation studies show node-specific patterns (NAPL) and data-driven graphs (DAGG) each contribute to improvements, with DAGG and unified embeddings yielding strong benefits.
• The model achieves over 5% relative improvements in MAE and MAPE over the strongest baselines on both datasets.
• DAGG analysis indicates the self-information term (identity) is important, and learned graphs resemble the properties of Chebyshev-based GCNs with similar performance to pre-defined graphs.
• Embedding dimension around 10 provides robust performance, balancing expressiveness and overfitting risk.