Source-linked AI summary
Spatio-Temporal Meta-Graph Learning for Traffic Forecasting
Renhe Jiang, Zhaonan Wang, Jiawei Yong, Puneet Jeph, Quanjun Chen, Yasumasa Kobayashi, Xuan Song, Shintaro Fukushima, Toyotaro Suzumura
TL;DR
Traffic forecasting must handle spatial and temporal heterogeneity, non-stationarity, and anomalous incidents. The paper introduces Meta-Graph Learning through MegaCRN, which uses a Meta-Node Bank within a GCRN encoder-decoder. MegaCRN outperformed state-of-the-art methods on METR-LA, PEMS-BAY, and EXPY-TKY while disentangling patterns and adapting to incidents.
Problem
Traffic forecasting models have not properly addressed spatio-temporal heterogeneity and non-stationarity, leaving different sensor signals entangled and incidents untreated.
Method
MegaCRN integrates a Meta-Graph Learner powered by a Meta-Node Bank into a GCRN encoder-decoder for spatio-temporal graph learning.
Results
MegaCRN outperformed state-of-the-art models on METR-LA, PEMS-BAY, and EXPY-TKY.
Takeaways & Limitations
Visualizations indicate that MegaCRN can disentangle road links and time slots with different patterns and adapt to incident situations.
Abstract
from arXiv · showhide
Traffic forecasting as a canonical task of multivariate time series forecasting has been a significant research topic in AI community. To address the spatio-temporal heterogeneity and non-stationarity implied in the traffic stream, in this study, we propose Spatio-Temporal Meta-Graph Learning as a novel Graph Structure Learning mechanism on spatio-temporal data. Specifically, we implement this idea into Meta-Graph Convolutional Recurrent Network (MegaCRN) by plugging the Meta-Graph Learner powered by a Meta-Node Bank into GCRN encoder-decoder. We conduct a comprehensive evaluation on two benchmark datasets (i.e., METR-LA and PEMS-BAY) and a new large-scale traffic speed dataset called EXPY-TKY that covers 1843 expressway road links in Tokyo. Our model outperformed the state-of-the-arts on all three datasets. Besides, through a series of qualitative evaluations, we demonstrate that our model can explicitly disentangle the road links and time slots with different patterns and be robustly adaptive to any anomalous traffic situations. Codes and datasets are available at https://github.com/deepkashiwa20/MegaCRN.
Introduction
Traffic forecasting combines spatial relationships among sensors with temporal patterns, but existing approaches do not properly address spatio-temporal heterogeneity, non-stationarity, and incidents. The paper proposes Meta-Graph Learning and MegaCRN to disentangle patterns and adapt to changing traffic situations.
- Motivation: Existing traffic forecasting models combine graph-based spatial modeling with recurrent, convolutional, or Transformer-based temporal modeling.These approaches represent spatial relationships with GCN-based modules and temporal dependencies with sequence models.
- Graph Structure Learning: Time-variant graph methods learn input-conditioned structures by projecting observations into node embeddings.The resulting spatio-temporal graph is described as a momentary graph.
- Research Gap: Spatio-temporal heterogeneity and non-stationarity remain insufficiently addressed, leaving different sensor signals entangled and incidents untreated.Attention mechanisms alleviate heterogeneity only to some extent.
- Proposed Approach: The proposed Meta-Graph Learner queries node-level prototypes from a Meta-Node Bank and reconstructs node embeddings with a Hyper-Network.This framework is designed to explicitly disentangle heterogeneity across space and time.
- Contributions: MegaCRN is evaluated quantitatively and qualitatively on METR-LA, PEMS-BAY, and the larger EXPY-TKY dataset with more complex incident situations.The authors state that the model is robust and adaptive from normal to non-stationary traffic situations.
Related Work
Traffic forecasting research progressed from statistical models to deep spatio-temporal architectures that combine graph operations with temporal modeling. Recent work increasingly learns graph structure adaptively or from observations.
- Traffic Forecasting: Early traffic forecasting used autoregressive, vector autoregressive, and autoregressive integrated moving average statistical models.
- Traffic Forecasting: Deep learning approaches subsequently became central to traffic forecasting and related multivariate time series forecasting.
- Temporal Modeling: Temporal Convolution and WaveNet provide long receptive fields for temporal modeling in several traffic forecasting architectures.
- Transformer Models: Traffic Transformers use self-attention-based architectures for spatio-temporal forecasting.
Problem Definition
The paper formulates traffic forecasting as multi-step-to-multi-step prediction over spatial units and information channels. Given α historical observations, the model infers the next β horizons.
- Forecasting Task: The forecasting model is trained to infer future observations from a fixed window of prior observations.The formulation is multi-step-to-multi-step rather than single-step prediction.
- Forecasting Task: The forecasting task maps previous α observation steps to the next β prediction horizons.The forecasting model F with parameters θ is trained to infer [X_t+1,...,X_t+β] from [X_t−(α−1),...,X_t].
- Data Representation: Each observation X_i has dimensions R^N×C, where N is the number of spatial units and C is the number of information channels.For this paper, the spatial units are road links and C equals 1 because only traffic speed is forecast.
Methodology
MegaCRN combines a GCRU encoder-decoder with a Meta-Graph Learner that generates situation-dependent graph structures from memory-augmented node embeddings. The framework uses a Meta-Node Bank to distinguish spatio-temporal traffic patterns and adapt graph topology across roads and time.
- MegaCRN: MegaCRN integrates a Meta-Graph Learner into a GCRU encoder-decoder for spatio-temporal modeling.The framework is presented as a generic architecture for learning graph structure from traffic observations.
- GCRU Encoder-Decoder: GCRU combines graph convolution over an input topology with recurrent processing to capture spatial and temporal dependencies.Its gates use graph convolution on the current observation and previous hidden state.
- Graph Structure Learning: Adaptive graphs derive topology from trainable node embeddings, whereas momentary graphs condition topology on observations or hidden states.Momentary graph variants may project hidden states or use self-attention before graph construction.
- Meta-Graph Learner: The Meta-Graph Learner uses a Hyper-Network and Meta-Node Bank to generate node embeddings for adaptive graph construction.The generated meta-graph replaces adaptive and momentary graphs as the auxiliary topology supplied to the GCRU encoder-decoder.
- Meta-Node Bank: The Meta-Node Bank stores typical features for pattern matching and reconstructs node representations from memory items matched to localized hidden-state queries.Each query is compared with memory items, and the resulting weighted combination augments the encoded hidden representation.
- Memory Regulation: Contrastive memory constraints treat the most similar prototype as positive and the second-most similar as negative, encouraging compact yet distinguishable prototypes.The authors report that adding these constraints to the MAE objective facilitates training convergence.
Experiment
MegaCRN is evaluated on standard traffic benchmarks and the large-scale EXPY-TKY dataset using forecasting comparisons, ablations, efficiency analysis, and qualitative studies. Results indicate strong accuracy, parameter efficiency, spatio-temporal disentanglement, and adaptation to traffic incidents.
- Quantitative evaluation: MegaCRN is compared with representative convolutional, recurrent, Transformer-based, adaptive-graph, and memory-network traffic forecasting baselines.The listed baselines include STGCN, DCRNN, GW-Net, STTN, GMAN, MTGNN, StemGNN, AGCRN, CCRNN, GTS, and PM-MemNet.
- Ablation study: The ablation study compares Adaptive GCRN, Memory GCRN, and Momentary GCRN variants to evaluate the contributions of MegaCRN components.The variants separately alter shared adaptive graphs, the Hyper-Network, and the Meta-Node Bank.
- Efficiency study: 133,597 parameters and the smallest overall MAE make MegaCRN the second-smallest model by parameter count in the EXPY-TKY efficiency comparison.The authors report comparatively efficient memory use and fast training rounds despite requiring more epochs to converge.
- Spatio-temporal disentanglement: Meta-graph embeddings cluster road links into groups whose shapes change over time, distinguishing rush-hour patterns from lower-speed, higher-variation interchange and toll-gate links.The clustering persists as time evolves, while the physical and daily-average patterns differ between the discovered groups.
- Incident awareness: During a traffic accident, MegaCRN better captures normal fluctuations and adapts to complex rush-hour and incident conditions than GW-Net and CCRNN.The local meta-graph also changes after the accident, with influence shifting away from road link 1 toward links 7–10.
Conclusion
MegaCRN combines a novel spatiotemporal graph structure learning mechanism with traffic forecasting and is evaluated across three datasets. It outperformed state-of-the-art models and qualitatively disentangled differing patterns while adapting to incident situations.
- MegaCRN introduces a novel spatiotemporal graph structure learning mechanism for traffic forecasting.
- MegaCRN was evaluated on METR-LA, PEMS-BAY, and the newly generated EXPY-TKY traffic dataset.EXPY-TKY was generated from large-scale car GPS records with corresponding traffic incident information.
- MegaCRN outperformed state-of-the-art models by a large degree on all three datasets.
- Visualizations demonstrated MegaCRN’s ability to disentangle time and nodes with different patterns and adapt to incident situations.