Dynamic Correlation Adjacency-Matrix-Based Graph Neural Networks for Traffic Flow Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Accurate traffic forecasting relies on modeling complex spatial and temporal dependencies in multivariate time series data.
• Graph convolutional networks (GCNs) show promise but are limited by static graph structures and difficulties in learning complex graph topologies.
• Existing methods may struggle to converge on datasets with challenging inherent graph structures.

Purpose of the Study:

• To propose a novel Dynamic Correlation Graph Convolutional Network (DCGCN) for enhanced multivariate time series forecasting.
• To address the limitations of predefined or static graph structures in GCNs for traffic prediction.
• To develop a model capable of dynamically learning spatial dependencies from data.

Main Methods:

• The proposed DCGCN constructs dynamic adjacency matrices using correlation coefficients to capture spatial dependencies.
• Gated temporal convolution is employed to effectively model temporal dependencies within the time series.
• The model integrates dynamic graph construction with temporal modeling for comprehensive dependency capture.

Main Results:

• Extensive experiments demonstrate the superior performance of DCGCN compared to ten baseline methods.
• The model achieved strong results on two original and four public datasets, validating its effectiveness.
• DCGCN successfully captured dynamic spatial dependencies crucial for accurate forecasting.

Conclusions:

• The Dynamic Correlation Graph Convolutional Network (DCGCN) offers a significant advancement in multivariate time series forecasting.
• Dynamic graph construction and gated temporal convolution effectively model complex dependencies in traffic data.
• DCGCN provides a robust and adaptable solution for traffic forecasting challenges.