[Paper Review] Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

Yaguang Li, Rose Yu+2 more|arXiv (Cornell University)|Jul 6, 2017

Neural Networks and Applications1,797 citations

TL;DR

The paper introduces DCRNN, a diffusion-convolutional recurrent network that models spatiotemporal traffic by diffusion on a directed graph, enabling improved long-term forecasting over baselines.

ABSTRACT

Spatiotemporal forecasting has various applications in neuroscience, climate and transportation domain. Traffic forecasting is one canonical example of such learning task. The task is challenging due to (1) complex spatial dependency on road networks, (2) non-linear temporal dynamics with changing road conditions and (3) inherent difficulty of long-term forecasting. To address these challenges, we propose to model the traffic flow as a diffusion process on a directed graph and introduce Diffusion Convolutional Recurrent Neural Network (DCRNN), a deep learning framework for traffic forecasting that incorporates both spatial and temporal dependency in the traffic flow. Specifically, DCRNN captures the spatial dependency using bidirectional random walks on the graph, and the temporal dependency using the encoder-decoder architecture with scheduled sampling. We evaluate the framework on two real-world large scale road network traffic datasets and observe consistent improvement of 12% - 15% over state-of-the-art baselines.

Motivation & Objective

• Model traffic speed forecasting on road networks as a spatiotemporal problem on a directed graph.
• Capture spatial dependencies via diffusion convolution on graph structures.
• Model temporal dynamics with recurrent units adapted to diffusion convolutions.
• Enhance long-term forecasting with an encoder-decoder setup and scheduled sampling.
• Empirically validate on large real-world datasets showing performance gains over baselines.

Proposed method

• Represent traffic sensors as a weighted directed graph and model diffusion via a random walk with restart to define diffusion convolution.
• Define a diffusion convolution operator that combines forward and reverse diffusion steps with trainable weights (K-step truncation).
• Implement a Diffusion Convolutional Layer to map input features to outputs using diffusion filters.
• Replace standard GRU/matrix multiplications with diffusion convolution to form DCGRU for spatiotemporal modeling.
• Use an encoder-decoder sequence-to-sequence framework with scheduled sampling to improve long-term forecasting.
• Train end-to-end to maximize likelihood of future time series on real-world datasets.

Experimental results

Research questions

• RQ1 How can diffusion processes on directed road graphs capture spatial dependencies for traffic signals?
• RQ2 Does diffusion-based spatial modeling combined with temporal recurrence improve short- and long-term traffic forecasts?
• RQ3 What is the impact of bidirectional diffusion versus unidirectional diffusion on forecasting accuracy?
• RQ4 Can encoder-decoder with scheduled sampling mitigate error propagation in multi-step forecasts?
• RQ5 How does DCRNN compare to spectral/undirected graph approaches in realistic traffic networks?

Key findings

• DCRNN outperforms state-of-the-art baselines across 15 min, 30 min, and 1 hour horizons on METR-LA and PEMS-BAY datasets.
• Bidirectional diffusion (forward and reverse) improves forecasting accuracy over unidirectional diffusion.
• Incorporating diffusion convolution significantly lowers validation error compared to models without spatial convolution.
• Encoder-decoder with scheduled sampling reduces error propagation in multi-step forecasting.
• DCRNN and its variants show stronger performance as forecast horizon increases, with notable gains on challenging datasets like METR-LA.