CoMemNet: A Dual-Branch Continual Learning Framework for Traffic Prediction

Researchers have introduced CoMemNet, a novel dual-branch continual learning framework designed to enhance traffic prediction in dynamic, non-Euclidean graph structures. Addressing the limitations of existing methods that rely on static graph structures, CoMemNet effectively captures continuously evolving patterns in streaming traffic networks. The architecture features a fast-converging Online branch for primary predictions and a momentum-updated Target branch that utilizes Wasserstein Distance features to create a Dynamic Contrastive Sampler. This sampler identifies nodes with significant dynamic changes, thereby mitigating catastrophic forgetting. Additionally, the system incorporates a lightweight Node-Adaptive Temporal Memory Buffer to consolidate historical knowledge while preventing memory explosion. The study presents two new open-source datasets and demonstrates that CoMemNet achieves state-of-the-art performance across three large-scale real-world datasets. The accompanying code has been made publicly available to support further research and application in intelligent transportation systems and machine learning.