Review of Urban Traffic Congestion Prediction Based on Multi-source Spatio-temporal Data Fusion and Graph Neural Network

Abstract

This review systematically explores the research progress of urban traffic congestion prediction based on multi-source spatio-temporal data fusion and graph neural networks (GNN). In response to the limitations of traditional methods in handling high-dimensional heterogeneous data and complex road network associations, the research focuses on three core challenges: the semantic alignment dilemma of multi-source data, the fragmentation of business decisions caused by the lack of model interpretability, and the high computing power cost of complex models that hinders actual deployment. By integrating cutting-edge literature, an interpretive lightweight alignment framework (ELAF) was proposed. Its innovation lies in three levels of technological breakthroughs: achieving cross-modal semantic interoperability by integrating dynamic graph construction and adversarial completion technology, constructing a human-machine collaborative decision-making mechanism by coupling attribution interpreters and rule extractors, and achieving lightweight deployment at the edge device level through deep reinforcement learning optimization. This framework successfully Bridges the triple gap of "data silos - algorithm black boxes - computing power bottlenecks". Empirical evidence shows that it has breakthrough significance in improving spatio-temporal consistency, enhancing decision understandability and optimizing resource efficiency. ELAF has for the first time established a technical closed loop of "data fusion - cognitive interaction - edge intelligence", providing real-time environmental cognitive capabilities for vehicle-road coordination systems, promoting a paradigm shift in traffic governance from passive mitigation to active intervention. This not only lays the theoretical foundation for the construction of intelligent transportation but also offers a systematic technical path for future research directions such as semantic perception enhancement and cross-domain transfer learning.