- 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.