TL;DR
XXLTraffic is a massive new open-source dataset for traffic forecasting that spans up to 23 years, dwarfing existing benchmarks in both duration and spatial scale. It introduces a novel "beyond test adaptation" framework, challenging models to predict traffic patterns across multi-year gaps to better simulate real-world urban evolution and infrastructure shifts.
What Problem Does This Paper Address?
Most current traffic forecasting research relies on datasets that are temporal "snapshots"—often covering only a few months or, at most, five years. While deep learning models have become excellent at capturing short-term spatio-temporal patterns, they struggle with the distribution shifts inherent in growing cities.
Over a decade, a city changes: new highways are built, populations surge, and vehicle usage patterns evolve. Furthermore, real-world sensors often suffer from long-term downtimes. Existing benchmarks don't account for these "extremely long" temporal gaps or the continuous evolution of the road network. The authors argue that we need to move "beyond test adaptation"—where a single model is tweaked to fit a new domain—toward models that can inherently generalize across the multifaceted impacts of demographic and vehicular growth over decades.
Methods
The researchers established a benchmarking framework focused on Extremely Long Prediction with Gaps. Unlike standard forecasting where the input and output are contiguous, this task introduces a gap ($g$) of 1, 1.5, or 2 years between the observation period and the prediction target.
They evaluated a wide array of architectures to see how they handle these massive shifts:
- General Time-Series Models: Informer, Autoformer, iTransformer, DLinear, and the state-space model Mamba.
- Traffic-Specific Models: STGCN, ASTGCN, GWN, AGCRN, and PDFor.
The benchmark includes both hourly and daily aggregated data, providing a multi-scale perspective on how temporal granularity affects long-term predictability.
Datasets
The paper introduces XXLTraffic, the largest public traffic dataset to date.
- Sources: Data was curated from the California Department of Transportation (PeMS) and Transport for NSW (Australia).
- Scale: It covers 9 districts in California and major roads in New South Wales, spanning from 2001 to 2024.
- Evolution: Unlike static datasets, XXLTraffic shows the network growing from a few hundred sensors to over 21,000 nodes over two decades.
Key Insights and Ideas
- Urban Evolution as Domain Shift: The authors demonstrate that sensors at the exact same location can have completely different data distributions in 2005 versus 2024. This confirms that long-term traffic forecasting is not just a time-series problem, but a domain adaptation problem.
- The "Aha" Moment for MICN: Interestingly, the MICN (Multi-scale Local and Global Context Network) model, which often underperforms on standard short-term benchmarks, emerged as a top performer in the "gap" setting. This suggests that models focusing on intrinsic data correlations and multi-scale patterns are more robust to long-term shifts than those relying on tight temporal dependencies.
- History Matters: The ablation studies showed that increasing the input sequence length (the "look-back" window) significantly improves performance even when predicting across a two-year gap. More historical context helps the model navigate the noise of long-term distribution shifts.
Key Results
- Performance Degradation: Nearly all SOTA models saw a significant drop in accuracy when faced with multi-year gaps, highlighting that current "leaderboard kings" are not yet ready for long-term urban planning applications.
- Efficiency Benchmarks: The researchers provided a comprehensive look at training costs. DLinear proved to be the most efficient (training in seconds), while graph-based models like AGCRN were computationally intensive, taking hours per epoch on the larger sub-datasets.
- Scale Invariance: Surprisingly, training on only 10% of the data yielded similar results to using the full dataset for some models, suggesting that the primary challenge is the difficulty of the task itself (the distribution shift) rather than a lack of data volume.
Limitations and Future Work
The primary limitation is the computational burden. The sheer size of 23 years of high-resolution traffic data makes it difficult for many researchers to train complex models without massive hardware resources.
The authors suggest that future work should focus on:
- Efficient Subsampling: Finding ways to represent 20 years of data without processing every single 5-minute interval.
- Foundation Models: Leveraging the scale of XXLTraffic to pre-train large-scale traffic foundation models that can be fine-tuned for specific cities or new road networks.
- Scalable Algorithms: Developing new architectures that can handle tens of thousands of evolving nodes over decades without hitting memory walls.
Summary updated