.png){kind=icon}
Return
EVOLVE: A Continual Learning Framework for Evolving Traffic Networks
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2026.104974.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Traffic flow prediction is central to intelligent transportation systems, yet existing spatial-temporal graph neural networks often assume static topology and stable traffic distributions, limiting their ability to handle evolving road networks. To this end, this study analyzes the spatial-temporal evolution characteristics of continuously evolving road networks using two California highway sub-networks, PeMS-D3 and PeMS-D4, over multi-year and multi-month periods. Through this analysis, we identify two patterns that motivate efficient continual learning within these datasets: spatial evolution is highly concentrated in a small fraction of nodes, and temporal series exhibit strong periodicity together with hierarchical information redundancy, with the first principal component contributing 77.3% of total variance on PeMS-D3 and 92.5% on PeMS-D4. Based on these insights, we propose the EVOLVE framework, which localizes evolving regions and constrains update scope through adaptive subgraph learning, reducing computational complexity; captures periodic patterns and compresses temporal features through dual-path temporal encoding, improving processing efficiency; and builds a parameter-pattern-data knowledge preservation system to mitigate catastrophic forgetting. Across the two evaluated benchmarks, EVOLVE achieves a 6.75% MAE improvement over the strongest baseline TEAM on PeMS-D3 and an 8.07% improvement on PeMS-D4, together with a 19 & times; training speedup and a 25 & times; GPU memory reduction relative to TEAM. These results support EVOLVE within the evaluated California highway networks, while generalization to non-highway and non-California networks remains an open direction for future work. The code of EVOLVE is available at github.com/OvOYu/EVOLVE.
Keywords:
Spatial-temporal data analytics
Graph neural networks
Continual learning
Evolving traffic networks
Journal
I
IF:
6.9
Papers:
549
Citations:
0
Organization
Cited Papers
No cited papers available