Abstract:Machine learning techniques for road networks hold the potential to facilitate many important transportation applications. Graph Convolutional Networks (GCNs) are neural networks that are capable of leveraging the structure of a road network by utilizing information of, e.g., adjacent road segments. While state-of-the-art GCNs target node classification tasks in social, citation, and biological networks, machine learning tasks in road networks differ substantially from such tasks. In road networks, prediction tasks concern edges representing road segments, and many tasks involve regression. In addition, road networks differ substantially from the networks assumed in the GCN literature in terms of the attribute information available and the network characteristics. Many implicit assumptions of GCNs do therefore not apply. We introduce the notion of Relational Fusion Network (RFN), a novel type of GCN designed specifically for machine learning on road networks. In particular, we propose methods that outperform state-of-the-art GCNs on both a road segment regression task and a road segment classification task by 32-40% and 21-24%, respectively. In addition, we provide experimental evidence of the short-comings of state-of-the-art GCNs in the context of road networks: unlike our method, they cannot effectively leverage the road network structure for road segment classification and fail to outperform a regular multi-layer perceptron.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: Existing graph convolutional networks (GCNs) have deficiencies when applied to road networks and cannot effectively use the structural information of road networks for prediction tasks. Specifically, existing GCNs are mainly designed for node classification tasks in social, citation, and biological networks, while tasks in road networks usually involve the prediction of edges (i.e., road segments), and many important tasks are regression tasks. In addition, road networks are very different from the networks assumed in existing GCN research in terms of attribute information and network characteristics, resulting in many implicit assumptions being invalid. ### Specific manifestations of the problem: 1. **Different task types**: - Existing GCNs mainly handle node classification tasks. - Tasks in road networks are usually prediction tasks on edges (such as speed estimation and speed limit classification), and many tasks are regression tasks. 2. **Different network characteristics**: - Road networks are edge - relational, containing not only node and edge attributes but also attributes that describe the relationships between road segments (such as angles). - Road networks have a lower density, and each road segment has fewer adjacent road segments, which makes neighborhood - aggregation - based methods sensitive to abnormal neighbors. - Homophily in road networks is fluctuating. There is high homophily within certain areas, but there may be sharp changes at the boundaries. ### Solutions: To solve these problems, the author introduced a new graph convolutional network - **Relational Fusion Network (RFN)**, specifically for machine - learning tasks in road networks. The main innovations of RFN include: 1. **Relational Fusion Operator**: - Extract useful information by aggregating relational representations instead of neighbor representations. - Use the attention mechanism to ignore the influence of abnormal neighbors. 2. **Multi - view learning**: - Consider both the node - relationship view and the edge - relationship view simultaneously to fully utilize nodes, edges, and inter - edge attributes. 3. **Improved aggregation function**: - Propose two fusion functions (AdditiveFuse and InteractionalFuse) to better capture fluctuating homophily. - Introduce an attention - mechanism - based aggregator that can filter out irrelevant or abnormal neighbors. ### Experimental results: The author conducted experiments on the road network in Aalborg, Denmark, and evaluated the performance of RFN in driving - speed - estimation and speed - limit - classification tasks. The results show that RFN significantly outperforms existing GCNs and other baseline methods in these tasks, improving performance by 32 - 40% and 21 - 24% respectively. In addition, standard GCNs failed to outperform multi - layer perceptrons (MLP) in these two tasks, indicating that they cannot effectively use neighborhood information. ### Summary: This paper aims to solve the limitations of existing GCNs in road - network applications by proposing RFN, especially in handling edge - prediction tasks and dealing with the unique structural characteristics of road networks. RFN significantly improves the performance of road - network prediction tasks by introducing the relational - fusion operator and the multi - view - learning framework.

Graph Convolutional Networks for Road Networks

Relational Fusion Networks: Graph Convolutional Networks for Road Networks

Attention Based Graph Covolution Networks for Intelligent Traffic Flow Analysis

Deep Graph Convolutional Networks for Accurate Automatic Road Network Selection

Road Network Representation Learning: A Dual Graph based Approach

Graph representation learning for road type classification

Road Extraction from Overhead Images with Graph Neural Networks

Traffic Graph Convolutional Recurrent Neural Network: A Deep Learning Framework for Network-Scale Traffic Learning and Forecasting

Holistically-Nested Structure-Aware Graph Neural Network for Road Extraction

Constructing Geographic and Long-term Temporal Graph for Traffic Forecasting

Graph learning-based spatial-temporal graph convolutional neural networks for traffic forecasting

Partitioned Graph Convolution Using Adversarial and Regression Networks for Road Travel Speed Prediction

GA-Net: A geometry prior assisted neural network for road extraction

On Representation Learning for Road Networks

Simpler is better: Multilevel Abstraction with Graph Convolutional Recurrent Neural Network Cells for Traffic Prediction

Informed along the road: roadway capacity driven graph convolution network for network-wide traffic prediction

Dual Dynamic Spatial-Temporal Graph Convolution Network for Traffic Prediction

Graph Pruning Based Spatial and Temporal Graph Convolutional Network with Transfer Learning for Traffic Prediction

3D Graph Convolutional Networks with Temporal Graphs: A Spatial Information Free Framework For Traffic Forecasting

Unboxing the graph: Neural Relational Inference for Mobility Prediction

Optimized Graph Convolution Recurrent Neural Network for Traffic Prediction