PMGCN: Progressive Multi-Graph Convolutional Network for Traffic Forecasting

Zhenxin Li, Yong Han, Zhenyu Xu, Zhihao Zhang, Zhixian Sun, Ge ChenISPRS Int. J. Geo Inf 2023 · DOI

TL;DR

This paper introduces PMGCN, a novel deep learning model that uses a progressive multi-graph convolutional network with spatiotemporal attention to improve traffic forecasting accuracy. By considering multiple graph structures (distance, similarity, and adaptive) and a progressive connection strategy, PMGCN captures complex and dynamic spatiotemporal dependencies more effectively than existing methods.

What Problem Does This Paper Address?

Accurate traffic forecasting is crucial for intelligent transportation systems, aiding in management, decision-making, and urban planning. However, traffic data exhibits complex, dynamic spatiotemporal dependencies that are challenging to model. Traditional methods often struggle with the non-Euclidean nature of road networks and the interplay between spatial relationships (which aren't solely based on distance) and temporal patterns (which aren't always periodic). Existing spatiotemporal graph neural networks typically rely on a single adjacency matrix, which can be insufficient to capture the full spectrum of spatial relationships. While some models incorporate multiple graphs or attention mechanisms, they often fail to dynamically extract comprehensive spatial correlations or effectively integrate information across different graph structures.

Methods

The core of PMGCN is its Progressive Multi-Graph Convolutional Network (PMGCN). It's built upon several key innovations:

  • Multi-Graph Construction: Instead of a single adjacency matrix, PMGCN constructs three distinct matrices to represent spatial relationships:
  • Distance Adjacency Matrix (AADJ): Based on the physical distance between road network nodes.
  • Similarity Adjacency Matrix (ASIM): Derived from the similarity of time series data between nodes, using Fast-DTW.
  • Adaptive Adjacency Matrix (AADA): Learned through a neural network, capturing complex, non-explicit relationships.
  • Progressive Connection: The three graph convolutional network (GCN) blocks are connected progressively. Each block processes information not captured by the previous ones, allowing for a step-by-step deep extraction of spatial correlations, starting from basic distance to more complex adaptive relationships.
  • Spatiotemporal Attention (STA): This module dynamically adjusts the importance of different spatial and temporal features. It incorporates:
  • Temporal Attention: Uses multi-head self-attention to capture long-range temporal dependencies.
  • Spatial Attention: Leverages the three adjacency matrices to refine attention scores, allowing for dynamic adjustment of spatial weights in graph convolutions.
  • Multi-Resolution Temporal Convolution (STC): A multi-scale 1D CNN with channel attention is used to extract temporal dependencies. This module employs an "Inception" like structure with filters of varying sizes (1x1, 1x3, 1x5) to capture correlations at different temporal scales and a lightweight channel attention mechanism (ECA) to focus on important channels.

Datasets

The paper uses the following datasets for training and evaluation:

  • Qingdao GPS Dataset: This dataset contains real GPS data from 340 trunk roads in Qingdao, China, providing traffic speed information sampled every 5 minutes.
  • PeMSD4 Dataset: A real California traffic flow dataset with 307 sensors, sampled from January to February 2018.
  • PeMSD8 Dataset: Another real California traffic flow dataset with 170 sensors, sampled from July to August 2016.

The datasets were split into training (60%), validation (20%), and testing (20%) sets, and normalized using z-score normalization.

Key Insights and Ideas

  • The Power of Multiple Perspectives: Using multiple, distinct adjacency matrices (distance, similarity, adaptive) is crucial for capturing the multifaceted nature of spatial relationships in traffic networks, going beyond simple proximity.
  • Progressive Learning for Depth: Connecting GCN blocks progressively allows the model to build upon previous layers' findings, ensuring that each layer extracts new, deeper spatial information rather than re-processing the same features. This avoids redundancy and enhances feature extraction.
  • Dynamic Attention is Key: Spatiotemporal attention, which dynamically adjusts feature importance based on context, is vital for handling the ever-changing nature of traffic conditions. The ability to adjust attention based on multiple graph structures further refines this.
  • Multi-Scale Temporal Analysis: Traffic patterns have dependencies at various time scales. The multi-resolution temporal convolution effectively captures these by using filters of different sizes, leading to more robust temporal modeling.

Key Results

PMGCN demonstrated superior performance across various traffic forecasting tasks:

  • Traffic Speed Forecasting (Qingdao GPS): PMGCN achieved the lowest Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) across different prediction horizons (15, 30, and 60 minutes). For example, at a 60-minute horizon, PMGCN reduced MAE by up to 20% compared to ARIMA and by 7% compared to the next best baseline (STGMN).
  • Traffic Flow Forecasting (PeMSD4 and PeMSD8): PMGCN also outperformed baseline models on these datasets. It reduced RMSE for PeMSD4 by up to 32% and for PeMSD8 by up to 33% at a 12-hour horizon.
  • Ablation Studies: Experiments showed that using only one type of graph (distance, similarity, or adaptive) resulted in worse performance than PMGCN's multi-graph approach. Removing spatial attention also led to a slight degradation in performance, highlighting the importance of each component.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited External Semantics: The model did not incorporate more comprehensive external semantic information (e.g., weather, points of interest, road types) which could further enhance predictions.
  • Limited Traffic Modes and Cities: The verification experiments primarily focused on traffic speed and flow, and were conducted in specific geographical areas. Further validation across different traffic modes and cities is needed.
  • Special Time Periods: The impact of specific time periods, such as morning and evening rush hours, was not deeply considered in the modeling.

Future work suggested includes:

  • Integrating relevant semantic information into the model.
  • Studying the model's adaptability to other traffic modes and cities.
  • Investigating the model's performance during special time periods and for more detailed prediction scenarios.
  • Further exploring and improving deep learning models for broader applications in traffic forecasting.

Summary updated

BibTeX
@article{li2023pmgcn,
  author = {Zhenxin Li and Yong Han and Zhenyu Xu and Zhihao Zhang and Zhixian Sun and Ge Chen},
  title = {{PMGCN: Progressive Multi-Graph Convolutional Network for Traffic Forecasting}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2023},
  doi = {10.3390/ijgi12060241},
}
Figure