acceptodds
Withdrawn from ICLR 2027

MASKED AUTOENCODER-BASED SPATIO-TEMPORAL GRAPH NEURAL NETWORK FOR ROAD TRAFFIC FLOW RECONSTRUCTION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traffic sensor networks frequently suffer from missing data due to equipment failures and communication disruptions, degrading downstream intelligent transportation system (ITS) applications that depend on complete observations. We propose MAST-GNN, a Masked Autoencoder-based Spatio-Temporal Graph Neural Network that reconstructs missing traffic measurements through co-optimized spatial and temporal components tailored to the reconstruction objective. A two-layer graph convolutional network propagates information from observed to masked nodes along the road topology. A multi-scale temporal encoder captures short-term fluctuations, daily periodicity, and long-term trends through three parallel branches fused via a learnable concatenation-projection. Seven structured masking strategies, four spatial and three temporal, simulate realistic sensor outage patterns. We evaluate MAST-GNN on production traffic data from 95 signalized intersections in Beijing, at two granularities that share a single navigation-derived source, an intersection-level flow view and a direction level turn-movement view, together with two public highway benchmarks. At a 30% mask rate, MAST-GNN attains R2 = 0.954 at the intersection level and R2 = 0.912 at the direction level, exceeding the strongest baseline at both granularities. Three paired ablations localize the gains: masked self-supervision (R2 0.061→0.682), multi-scale temporal encoding (0.682→0.912), and graph propagation under cold start (0.886 versus −0.024 for non-parametric spatial interpolation). A cold-start study on twenty intersections with no training history exposes the complementary roles of the two information sources: temporal periodicity dominates under normal operation, whereas spatial topology becomes the sole recovery channel when temporal information fails. A mask-rate sweep quantifies this conditional value: the gain over the time-only baseline remains negligible under normal operation (−0.001 at 30% masking) and grows with temporal deprivation, reaching +0.224 at 90% masking; under cold start, graph propagation restores R2 = 0.886 versus −0.024 for non-parametric spatial interpolation (+0.910). On public benchmarks, MAST-GNN achieves R2 = 0.731 on METR-LA (207 highway sensors) and R2 = 0.886 on PEMS-BAY (325 Bay Area sensors), demonstrating robust transfer to real-world highway networks. Cross-dataset analysis reveals that model capacity, graph connectivity, and feature dimensionality interact multiplicatively. The hidden dimension must exceed a graph-complexity-dependent threshold before auxiliary features can contribute substantially to reconstruction accuracy.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Spatio-Temporal Graph Neural Network,MASKED AUTOENCODER,TRAFFIC FLOW RECONSTRUCTION

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9oxNOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method called MAST-GNN for filling in the missing traffic data in the signalized road network. This method employs a two-layer GCN, first propagating zero-filled observations on the graph. Subsequently, three branches, namely short-term FC, sinusoidal periodic encoding and global mean pooling, are fused and processed through a one-dimensional convolution with a kernel size of 7 and a three-layer fully connected decoder. The training process combines four spatial structured masks and three temporal structured masks and uses a weighted MSE loss, with weights determined by peak hours and node variance. The authors demonstrated through cold-start tests and repeated retraining experiments with different mask rates that, in normal operation, the graph structure is almost redundant and only becomes the primary recovery channel when temporal information is lost.

Critical strengths

The key strengths of this paper are that a single navigation source simultaneously provides intersection-level and twelve-movement direction-level views, along with signal timing information. The left-turn R² is approximately 0.81, the through-movement R² is 0.92, and the right-turn flow is non-zero only 20% to 28% of the time. This makes the testbed closer to the actual deployment scenario than the usual public highway benchmarks.

Critical weaknesses

W1. The key claim is that the graph becomes the only recovery channel once temporal information is missing. However, the only cold-start comparator in Tables 6 and 2 is spatial KNN, which actually uses graph neighbors yet sits in the graph-free column of Table 2. Without MAST-GNN minus the GCN or a same-slot mean across intersections, we can’t tell how much of the 0.886 comes from the graph, and the main claims in the abstract, Contribution 3 and the conclusion all hinge on this result.

W2. The gains of 0.912 over 0.910 for T-CNN and 0.954 over 0.946 for SAITS are single-seed results, while the authors themselves report about ±0.002 run-to-run variation. In the 30% row of Table 2, T-CNN actually wins 0.911 to 0.910. SPIN, STD-MAE, MagiNet, GRIN [1] and ImputeFormer [2] are not compared, and Table 1 has no baselines at all. As a result, neither the claim of beating the strongest baseline nor the claim of robust transfer is convincingly supported.

References

[1] Cini A, Marisca I, Alippi C. Filling the g_ap_s: Multivariate time series imputation by graph neural networks[J]. arXiv preprint arXiv:2108.00298, 2021.

[2] Nie T, Qin G, Ma W, et al. ImputeFormer: Low rankness-induced transformers for generalizable spatiotemporal imputation[C]//Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 2024: 2260-2271.

Questions for authors

Q1. In the cold-start experiment, do the training-period ground truths of the 20 intersections go into the loss? If so, could you report R² in two settings, one with these nodes fully excluded from the loss, and one with the GCN removed and only the temporal encoder kept?

Q2. Is the 30% mask in Tables 2, 4 and 5 element-wise over time and intersection, or the column-wise spatial plus row-wise temporal masks from Appendix B.13? If it’s the latter, how does a per-node T-CNN recover a fully masked column?

Sug1. Comparing on more benchmark datasets would make the results more convincing.

Sug2. The writing has quite a few errors. GMAN is cited as Song et al. 2020, which is actually STSGCN. PCT is described as masked point cloud modeling, and T-GCN is described as temporal convolution even though it uses GRU. The cold-start result of 0.886 vs. -0.024 also shows up twice in the abstract, and the pile of numbers hurts readability.

Reviewer_Bfs2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on the missing data in spatiotemporal graphs and proposes a reconstruction task. However, neither the problem nor the method is novel. There has been a line of research on spatiotemporal Kriging (interpolation) [1]; and masked autoencoder and GNNs are well established. There seems a new intersection data from Beijing, but at a limited scale (95 nodes). Baseline models used in the experiments are all published several years ago...

[1] Inductive graph neural networks for spatiotemporal kriging. AAAI 2021

Critical strengths

None

Critical weaknesses
  • W1: Trivial problem (reconstruction) that has been resolved by an existing line of research, namely spatiotemporal Kriging.
  • W2: No novelty, by leveraging well-established modules, e.g., masked autoencoder, GNNs.
  • W3: Baselines are naive and old, published before 2023.
Questions for authors

None

Reviewer_xtuJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MAST-GNN, a model that integrates GCNs, multi-scale temporal encoding, and masked self-supervised learning to reconstruct missing total intersection traffic volumes and turning movement flows. The model incorporates seven distinct masking strategies to simulate various patterns of data loss and has been validated on both public and private datasets. Experimental results demonstrate that while temporal models perform well under typical data loss conditions, the inclusion of spatial information becomes significantly more beneficial when data loss is severe.

Critical strengths

S1. The research problem has practical significance. The paper addresses missing traffic data by reconstructing both total intersection traffic flow and turning flow, which is a task with clear applications in traffic management.

S2. The methodology is intuitive. The design components of the model are logically linked to the imputation task. Each component is logically linked to the imputation task, and the inclusion of seven distinct masking strategies demonstrates the authors' careful consideration of various data-loss scenarios.

S3. The experimental analysis goes beyond simple comparisons of average performance by exploring the conditions under which the model is most effective. The paper examines scenarios involving varying missing-data rates, different levels of data granularity, and cold-start conditions, while also offering insights into the respective roles of temporal and spatial information under these diverse circumstances.

Critical weaknesses
  1. The description of key experimental details is insufficient, making it difficult to verify the conclusions regarding "cold-start" scenarios and the scope of information usage. Specifically, the cold-start experiment fails to clearly distinguish between the masking of inputs and the exclusion of data from the training supervision process. In lines 399–403, the authors simulate a "no historical data" cold-start scenario by masking 20 intersections throughout both the training and testing phases. However, lines 274–278 and Equation (13) indicate that the ground-truth values for these masked locations are still utilized in the reconstruction loss calculation. The paper does not clarify whether these cold-start nodes are also excluded from the training supervision set. If the ground-truth historical data for these nodes remains part of the loss computation, the experiment cannot adequately support the claim that the model addresses cold-start scenarios involving "new intersections for which no historical data has ever been provided."

  2. The baseline coverage is insufficient, failing to sufficiently establish the competitiveness of the proposed method against representative spatial-temporal imputation models. Comparisons with other baselines (currently in Tables 4 and 5) should be moved to the main body of the paper. Furthermore, the study lacks comparisons with other prominent spatial-temporal imputation methods from recent literature—such as ImputeFormer (KDD 2024) and PriSTI (ICDE 2023), which represent Transformer-based and diffusion-based approaches, respectively.

Analyzing the role of graph structure and evaluating the overall competitiveness of the proposed method are distinct tasks; the former cannot substitute for the latter. Even if the primary focus of the paper is the value of graph structure, it should compare the method against established, representative imputation models. Additionally, Table 1 only reports the performance of MAST-GNN under various masking strategies, further limiting the ability to assess its relative performance against public benchmarks.

  1. The paper lacks validation on publicly available traffic flow datasets. Although the authors perform total-flow and turning-flow reconstruction using a private dataset from Beijing, both views originate from the same data source and cannot be considered independent datasets. It is recommended that the authors incorporate several representative public traffic flow datasets—such as PEMS03, PEMS04, PEMS07, or PEMS08.

  2. There is an inconsistency between the figures and the main text: Figure 1 is labeled with "6 Masking Strategies" and "Temporal Convolution (kernel=3)," whereas the text specifies seven masking strategies and a convolution kernel size of 7.

Reference:

Liu, Mingzhe, Han Huang, Hao Feng, Leilei Sun, Bowen Du, and Yanjie Fu. "Pristi: A conditional diffusion framework for spatiotemporal imputation." In 2023 IEEE 39th international conference on data engineering (ICDE), pp. 1927-1939. IEEE, 2023.

Nie, Tong, Guoyang Qin, Wei Ma, Yuewen Mei, and Jian Sun. "ImputeFormer: Low rankness-induced transformers for generalizable spatiotemporal imputation." In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pp. 2260-2271. 2024.

Questions for authors

Please refer to the weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.