DDMGPN: A derivative-driven multi-graph propagation network with traffic knowledge graph for traffic flow prediction

Abstract

In dynamic urban environments, accurate traffic flow prediction faces three major challenges: intricate spatio-temporal dependencies, integration of heterogeneous data, and abrupt state changes. This paper proposes a novel Derivative-Driven Multi-Graph Propagation Network (DDMGPN), synergized with a Traffic Knowledge Graph (TKG) to address these challenges. The TKG integrates multi-source data (e.g., road topology, points of interest, overhead view images) into a unified knowledge representation to systematically encode prior knowledge. Building upon this foundation, DDMGPN introduces three innovative components to enhance spatio-temporal modeling. First, a derivative-driven feature modulation mechanism integrates first and second derivatives of traffic flow data, enabling joint modeling of trend evolution and abrupt state changes in traffic flow. Second, a multi-graph synergistic architecture combines a knowledge-guided static prior graph, a flow evolution dynamic graph, and a flow variation dynamic graph, establishing a three-stage knowledge propagation paradigm for spatio-temporal modeling. Finally, a temporal propagation amplifier (TPA) incorporates adaptive attention and derivative amplification, mitigating error accumulation in multi-step predictions. Comprehensive experimental evaluations conducted on two real-world datasets show that DDMGPN achieves state-of-the-art performance, both for short-term predictions and long-term predictions. Moreover, we visualize the learned spatio-temporal adjacency matrix to enhance the interpretability of our proposed model.