Continuous Cross-Domain Traffic State Prediction via Memory-Augmented Graph Liquid Time-Constant Networks

The paper presents the Memory-Augmented Graph Liquid Time-Constant Network (MA-GLTC), a novel framework for continuous cross-domain traffic state prediction. It introduces spatio-temporal units for fine-grained knowledge alignment and a graph liquid time-constant network that enhances traffic evolution modeling through adaptive time constants and neighborhood feedback. The experimental results on five public datasets show that MA-GLTC outperforms existing methods, achieving an average prediction error reduction of up to 10.09%, highlighting its potential for improving traffic prediction in data-scarce environments.