Continuous Cross-Domain Traffic State Prediction via Memory-Augmented Graph Liquid Time-Constant Networks
Researchers have developed a new framework called Memory-Augmented Graph Liquid Time-Constant Network (MA-GLTC) for predicting traffic states across different domains. This method addresses limitations in existing approaches by enabling fine-grained adaptation between source and target domains, better handling of unseen patterns, and more accurate modeling of continuous traffic dynamics. MA-GLTC utilizes spatio-temporal units for knowledge alignment and a graph liquid time-constant network with a memory mechanism to preserve and update traffic knowledge, demonstrating superior performance over baseline methods in various prediction tasks. AI