Authors
James Jianqiao Yu

Publication
IEEE Transactions on Knowledge and Data Engineering, in press, 2024

Abstract
Graph-based deep learning models are becoming prevalent for data-driven traffic prediction in the past years, due to their competence in exploiting the non-Euclidean spatial-temporal traffic data. Nonetheless, these models are approaching a limit where drastically increasing model complexity in terms of trainable parameters cannot notably improve the prediction accuracy. Furthermore, the diversity of transportation networks requires traffic predictors to be scalable to various data sizes and quantities, and ever-changing traffic dynamics also call for capacity sustainability. To this end, we propose a novel adaptive deep learning scheme for boosting graph-based traffic predictor performance. The proposed scheme utilizes domain knowledge to decompose the traffic prediction task into sub-tasks, each of which is handled by deep models with low complexity and training difficulty. Further, a stream learning algorithm based on the empirical Fisher information loss is devised to enable predictors to incrementally learn from new data without re-training from scratch. Comprehensive case studies on five real-world traffic datasets indicate outstanding performance improvement of the proposed scheme when equipped to six state-of-the-art predictors. Additionally, the scheme also provides impressive autoregressive long-term predictions and incremental learning efficacy with traffic data streams.