- Rating
- 3(+2)
- Confidence
- 4
- Informativeness
- 3(+1)
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes PIANO, a graph neural ODE approach to learn on graph-structured data. For spatial propagation, PIANO considers a learnable convex combination of different-length updates, making it multi-hop and able to go beyond 1-WL tests and MPNN architectures. Node states evolve following a non-linear ODE based on this combination of spatial updates. The flow of the model is well-posed, and the node- and topology-dependent gate is sound for distinguishing topologies around nodes. PIANO is tested on both spatiotemporal and static tasks, with an initial validation of the mechanisms in more synthetic settings.
Critical strengths
-
The paper is well structured and motivated. The theory section is well written and provides useful statements on the motivation and well-posedness of the model, while also highlighting its advantages and how it can distinguish local geometries around nodes.
-
The paper provides enough ablations, technical and implementation details to understand the components of the model and how they work together.
Critical weaknesses
The paper fails to reference or compare with a long list of relevant methods in the field, more specifically, in two categories:
- Multi-hop and Path GNNs. Since PIANO essentially weights paths and powers of the adjacency matrix, it should be compared to other state-of-the-art models with similar architectures, such as MixHop [1], DReW [2], PathNN [3], and PANConv [4] (needs a deeper comparison, as it is based on Path-integral as well).
- Differential equation-inspired and Graph ODE models. Similarly, a lot of models are based on latent ODEs and have different propagation guarantees depending on the adopted dynamics, such as PDE-GCN [5], GraphCON [6], ADGN [7], PH-DGN [8], SWAN [9], and SONAR [10].
Furthermore, the authors could compare the models to well-known benchmarks for spatiotemporal propagation (PEMS-Bay and Metr-LA [11], EngRad [12]) or static data (such as LRGB [13]), where the multi-hop and continuous-time propagation dynamics have shown advantages.
Finally, the controlled experiments lack details as of now, especially in the main text. The authors should consider adding more details and interpretation. As of now, it is hard to understand the setting and especially the interpretation of the results.
[1] Abu-el-Haija et al., MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing, ICML 2019.
[2] Gutteridge et al., DRew: Dynamically Rewired Message Passing with Delay, ICML 2023.
[3] Michel et al., Path Neural Networks: Expressive and Accurate Graph Neural Networks, ICML 2023.
[4] Ma et al., "Path Integral Based Convolution and Pooling for Graph Neural Networks", NeurIPS 2020
[4] Eliasof et al., PDE-GCN: Novel Architectures for Graph Neural Networks Motivated by Partial Differential Equations, NeurIPS 2021.
[5] Rusch et al., Graph-Coupled Oscillator Networks, ICML 2022
[6] Gravina et al., Anti-Symmetric DGN: A Stable Architecture for Deep Graph Networks, ICLR 2023.
[7] Heilig et al., Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks, ICLR 2025.
[8] Gravina et al., On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems, AAAI 2025.,
[9] Trenta et al., SONAR: Long-range propagation through Information waves
[10] Li et al., Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. ICLR 2018
[11] Marisca et al., Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling. ICLR 2024.
[12] Dwivedi et al., Long Range Graph Benchmark, NeurIPS 2022
Questions for authors
- It seems like the bolding in some tables is wrong or misleading (for example, in Table 2 or Table 3, where PIANO is bold but not the best).
- Is it possible to extend Theorem 2 to a non-linear model? What are the main differences and advantages?
- Could the authors compare to other multi-hop and path GNNs, showing advantages and/or main differences?
- The authors should add more details on the controlled experiments showing the effects of the mechanisms.
11 comments
CW2 / Q3 — Weather and traffic comparisons
Main-paper results: transfer across tasks
Tables 4–5 give PIANO the best displayed means on Ocean (0.5245 ± 0.0013 Mean ACC; 43.1K parameters), SEVIR (all three metrics; 69.9K parameters) and PubMed (89.645 ± 0.534%). Ocean uses PIANO's separately selected recipe alongside original baseline recipes; B.4/E.6 disclose their different budgets.
EngRAD: forecasting across weather stations
EngRAD (Marisca et al., 2024) uses 487 locations, five channels, 24 observed/24 predicted hours and chronological 2018/2019/2020 splits, with train-only normalization. Mean ± sample SD over ten seeds:
| Method | Params | Std. MAE ↓ | Std. MSE ↓ |
|---|---|---|---|
| MixHop | 55.9K | 0.4252 ± 0.0011 | 0.5231 ± 0.0027 |
| PAN convolution | 55.9K | 0.4268 ± 0.0020 | 0.5246 ± 0.0027 |
| DReW | 89.1K | 0.4289 ± 0.0015 | 0.5311 ± 0.0066 |
| PathNN-SP | 80.8K | 0.4285 ± 0.0013 | 0.5269 ± 0.0038 |
| PDE-GCN | 55.7K | 0.4225 ± 0.0011 | 0.5173 ± 0.0025 |
| GraphCON | 55.9K | 0.4262 ± 0.0008 | 0.5250 ± 0.0043 |
| A-DGN | 55.8K | 0.4249 ± 0.0013 | 0.5215 ± 0.0029 |
| PH-DGN-C | 51.7K | 0.4273 ± 0.0011 | 0.5242 ± 0.0023 |
| SWAN | 59.9K | 0.4252 ± 0.0014 | 0.5225 ± 0.0033 |
| SONAR | 80.8K | 0.4253 ± 0.0010 | 0.5208 ± 0.0018 |
| GRAND | 55.7K | 0.4224 ± 0.0012 | 0.5169 ± 0.0023 |
| GREAD-BS | 47.5K | 0.4240 ± 0.0012 | 0.5186 ± 0.0026 |
| PIANO | 56.2K | 0.4228 ± 0.0014 | 0.5197 ± 0.0027 |
PIANO's MAE is close to GRAND/PDE-GCN and below the other displayed architectures, including MixHop, DReW and PathNN-SP. This demonstrates competitive transfer of explicit spatial-order control to a new weather environment.
PEMS-BAY: accuracy and model size on a road graph
PEMS-BAY (Li et al., 2018) predicts twelve five-minute steps from twelve observations on a road graph, with chronological 70/10/20 splits and train-only normalization. MAE is in native units; Mean MAE pools all twelve leads.
| Method | Params | MAE₁₅ ↓ | MAE₃₀ ↓ | MAE₆₀ ↓ | Mean MAE ↓ |
|---|---|---|---|---|---|
| MixHop | 36.4K | 1.383 ± 0.001 | 1.813 ± 0.001 | 2.345 ± 0.003 | 1.778 ± 0.001 |
| PAN convolution | 36.4K | 2.505 ± 0.389 | 2.755 ± 0.311 | 3.139 ± 0.229 | 2.764 ± 0.323 |
| DReW | 136.3K | 1.352 ± 0.003 | 1.730 ± 0.004 | 2.140 ± 0.005 | 1.679 ± 0.003 |
| PDE-GCN | 36.2K | 1.402 ± 0.001 | 1.879 ± 0.002 | 2.512 ± 0.004 | 1.856 ± 0.002 |
| GraphCON | 28.1K | 1.391 ± 0.001 | 1.863 ± 0.002 | 2.481 ± 0.004 | 1.838 ± 0.002 |
| A-DGN | 28.0K | 1.387 ± 0.001 | 1.840 ± 0.001 | 2.413 ± 0.004 | 1.809 ± 0.001 |
| PH-DGN-C | 23.9K | 1.420 ± 0.015 | 1.898 ± 0.028 | 2.523 ± 0.057 | 1.872 ± 0.032 |
| SWAN | 32.1K | 1.388 ± 0.001 | 1.838 ± 0.001 | 2.415 ± 0.003 | 1.808 ± 0.002 |
| SONAR* | 53.0K | 1.410 ± 0.001 | 1.894 ± 0.004 | 2.546 ± 0.010 | 1.874 ± 0.005 |
| GRAND | 28.0K | 1.432 ± 0.002 | 1.937 ± 0.004 | 2.635 ± 0.009 | 1.922 ± 0.004 |
| GREAD-BS | 19.8K | 1.437 ± 0.000 | 1.950 ± 0.001 | 2.652 ± 0.002 | 1.933 ± 0.001 |
| PIANO | 28.4K | ± 0.001 | ± 0.003 | ± 0.006 | ± 0.003 |
Bold marks the best displayed means; underlining marks second-best means. SONAR* uses five seeds at the paper's reporting cutoff; other rows use ten. Protocols and completion details: B.4; I.3; I.5.
PIANO offers a strong accuracy–capacity trade-off. Its separately validation-selected setting achieves 1.7583 ± 0.0035 Mean MAE, second among the completed ten-seed methods listed and below MixHop's 1.7782 ± 0.0014. PIANO uses 28.4K parameters, only 20.9% of DReW's 136.3K; DReW achieves 1.6794 ± 0.0028. The table retains original PIANO; the additional selection budget is disclosed in I.3.
METR-LA: transfer to a second traffic dataset
PIANO's separately selected setting achieves MAE 3.4473 ± 0.0049, below fixed-recipe GRAND/GREAD-BS 3.7455 ± 0.0028 / 3.7610 ± 0.0068. DReW and the completed ten-seed PathNN-SP give 3.2860 ± 0.0118 / 3.3085 ± 0.0158, respectively (Table 41).
Ocean: measured computational efficiency
PIANO uses 43.1K versus PathNN-SP's 67.7K parameters. The isolated audit measures 0.790 versus 11.763 s/epoch and 0.415 versus 6.976 ms per nominal sample (Table 28). These timings use original LR0.001 checkpoints; the selected LR0.003 accuracy setting is separate. Historical ConvLSTM reaches higher overall Mean ACC, 0.547 ± 0.004; full cohorts are in Appendix E.
Together, PIANO combines competitive forecasting with a distinct architectural contribution: efficient redistribution of propagation mass across explicit spatial orders as the latent state evolves within a continuous-time graph model.
Q1, Q2 & Q4 — Highlighting, nonlinear theory and controlled experiments
Q1 — Corrected table highlighting
We corrected bolding and clarified the comparison sets. Table 2 now marks CSG-ODE as best on Wave (3.26 versus PIANO's 3.79, in MSE). Full reference sets, including Graph WaveNet and ConvLSTM, remain in Appendices E–F.
PIANO has the best displayed main-table means on Ocean, SEVIR and PubMed, and is competitive with the strongest EngRAD baselines. Together with compact models and measured Ocean efficiency, these results support a favorable accuracy–resource trade-off (Comment 2).
Q2 — Scope of the nonlinear theory
Theorem 2 bounds frozen linear transport, isolating the roles of spatial order and integration horizon . In the full model, derivatives also pass through the globally pooled controller, adding coefficient sensitivity beyond local feature transport. Extending the radius analysis requires separate bounds on these effects. Theorem 1 already establishes well-posedness of the full nonlinear flow, whose spatial-order weights adapt to the evolving state. Revised §4.2 and Appendix C clarify this distinction.
CW3 / Q4 — Controlled settings and interpretation
We expanded §5.1 and Appendix D, with direct links to Figures 2 and 4.
Wave—adaptive propagation range. A fixed 64-node ring connects first/second neighbors. With the gate off, models observe ten displacement/velocity frames and predict twenty; hidden varies interaction range. Learned effective radius tracks (Spearman 0.943 over 200 matched families for the diagnostic checkpoint). Swapping only walk-length schedules at matched solver times increases error by orders of magnitude, demonstrating functional use of the learned weights.
Springs—structural modulation. Ten particles use the same 10-to-20-frame task, with a supplied graph resampled per trajectory and stiffness or . At weak coupling, removing/shuffling structural information increases MSE from 5.28 × 10⁻⁴ to 8.85–9.12 × 10⁻⁴. Effects are small at strong coupling. This identifies a useful regime for correctly aligned structural information and motivates the gate's optional, task-dependent role.
Q4 — Matched controllers and ongoing feedback
-frozen fixes weights predicted from the initial state; -conditioned uses the initial state and solver time; full PIANO uses the evolving state and time. Within each task, architecture, parameter count and field-evaluation count are matched. Results are mean ± sample SD over ten paired seeds (Table 3; D.11):
| Controller | Source-disjoint Wave: MSE ↓ | Bounded SEVIR: MSE ↓ | Ocean: Mean ACC ↑ |
|---|---|---|---|
| -frozen | 4.087 ± 0.686 | 6.6199 ± 0.1002 | 0.523110 ± 0.000975 |
| -conditioned | 4.144 ± 0.823 | 6.6147 ± 0.0622 | 0.523152 ± 0.000969 |
| Full evolving-state PIANO | 3.874 ± 0.620 | 6.6036 ± 0.0572 | 0.524505 ± 0.001302 |
Full has the best displayed mean in all three settings. Ocean improves in all ten pairs against each control, supporting a practical benefit from ongoing feedback. Wave/SEVIR mean improvements remain uncertain in the paired analysis. Wave uses the source-disjoint companion; bounded SEVIR uses a separate event-disjoint protocol.
CW1 / Q3 — Related work and architectural differences
Thank you for identifying the missing comparisons and the points needing clarification. We now cite and evaluate all ten requested model families and add EngRAD, METR-LA, PEMS-BAY and Peptides-func comparisons. We also revised the operator comparison, clarified nonlinear theory, and added controlled-experiment explanations. Substantive paper revisions are orange. This response gives the contribution/related work first, forecasting comparisons in Comment 2, direct theory/presentation answers in Comment 3, and LRGB/additional graph results in Comment 4.
Our Central question is: How can a continuous-time graph model redistribute propagation mass across spatial orders as its latent state evolves, while retaining continuous temporal evolution and affordable explicit multi-order propagation?
PIANO makes spatial order an adaptive control variable inside each vector-field evaluation: Mθ(H,t)=∑k=0Lαk(H,t)A^k, with α(H,t) on the simplex. Sparse recursion costs O(L∣E∣d) for propagation. This combines explicit order control with continuous-time evolution.
Requested multi-hop/path methods. We now cite and compare MixHop (Abu-El-Haija et al., 2019), DReW (Gutteridge et al., 2023), PathNN (Michel et al., 2023) and PAN (Ma et al., 2020). Appendix H explains the distinctions:
Deeper comparison with PAN. PAN learns layer-level path-length weights and uses the operator diagonal for structural pooling. PIANO recomputes its distribution from the evolving state and solver time. Our optional gate shares the closed-walk motivation but modulates node-wise field updates. Our PAN baseline is explicitly a node-preserving convolution-only adaptation, without PANPool; B.3–B.4/H identify this choice.
Requested differential-equation models. Revised Table 1 and H.2 add PDE-GCN (Eliasof et al., 2021), GraphCON (Rusch et al., 2022), A-DGN (Gravina et al., 2023), PH-DGN (Heilig et al., 2025), SWAN (Gravina et al., 2025) and SONAR (Trenta et al., 2025). Their diffusion, oscillator, antisymmetric, Hamiltonian and wave dynamics provide complementary propagation mechanisms. Table 1 now distinguishes explicit spatial reach from temporal derivative order and reach accumulated through integration. It recognizes PH-DGN's possible two-hop energy-gradient reach and GREAD-BS's two-hop reaction.
PIANO's distinction is the state-conditioned redistribution among explicit spatial orders within the continuous vector field.