acceptodds
Under review as a conference paper at ICLR 2027

Gaussian Graph(ical) Neural Networks for repeated observations over a fixed graph with persistent edge relationships

Abstract

Many networked systems are observed repeatedly over a fixed physical or relational structure. In applications such as environmental monitoring, individual observations may be sparse, noisy, or incomplete, making the current state difficult to estimate from measurements alone. However, repeated observations provide an opportunity to learn dependencies between nodes that persist across measurements and can constrain future inference. Existing graph neural networks do not naturally encode such dependencies, while classical Gaussian graphical models encode them explicitly but offer limited representational flexibility. We introduce Gaussian Graphical Neural Networks, which combine expressive neural representations with Gaussian graphical inference by learning a sparse precision matrix over the fixed graph. We develop three precision-matrix parameterizations that guarantee convergent inference, characterize the spectral behavior of the resulting operator and show that it imposes no inherent smoothing bias, and derive an implicit differentiation procedure that enables end-to-end training without unrolling inference. Experiments on spatial reconstruction and hydrological forecasting demonstrate that learning persistent graphical dependencies and performing structured inference over latent representations can improve estimation in repeated-observation settings.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Gaussian Graphical Model, Gaussian Belief Propagation, Precision Matrices, Implicit Differentiation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_QS9WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a Gaussian Graphical Neural Networks which learns a sparse precision matrix J at each layer and computes the µ = J⁻¹h with Gaussian belief propagation. Three parameterizations (pairwise-normalizable, diagonally dominant, Laplacian) guarantee walk-summability and hence convergence. And experiments cover temperature imputation on USCRN (131 stations) and streamflow forecasting on LamaH-CE (608 gauges), where GGNN outperforms GCN-style baselines and is competitive with GraphGPS and Transformers.

Critical strengths

1.The method is well-motivated and somehow control the training memory.

2.The parameterizations guarantee convergent inference by construction.

3.The paper is clearly written and well organized, and includes useful analyses including an iteration-budget ablation and visualizations of the learned precision matrices.

Critical weaknesses

1.The baselines are not fair and completed .Without positional encodings, GCN, IGNN and GPR-GNN cannot distinguish nodes (MAE ≈ 8 on USCRN, worse than a graph-free Transformer). The no-PE gains may therefore reflect node-specific parameters rather than graphical inference. Only GCN-E has learnable edge weights. Established kriging and spatiotemporal baselines are missing (IGNNK, GRIN, Graph WaveNet, AGCRN), as is Deep GMRF.

2.The datasets is construct by the setting only in this paper. The USCRN graph and masks, and the LamaH-CE subset, inputs and horizons, are all defined by the authors. No result can be compared with published numbers. Prior work on LamaH-CE (Kirschstein & Sun, 2024) found that river topology does not help GNNs, but it is neither cited nor addressed.

3.The novelty of this paper is a little limited and the scalability of this method is not clear. Each layer is a linear filter J⁻¹ with learned edge weights, close to existing implicit and diffusion GNNs. The Gaussian semantics are unused: there is no uncertainty output. On graphs this small, a direct solver or conjugate gradient would only require positive definiteness, which calls the walk-summability machinery into question. Training is about 5.5× slower than GCN, and scalability beyond 608 nodes is not demonstrated.

Questions for authors

1.Could you evaluate on at least one established benchmark under its published protocol, e.g. METR-LA or PEMS-BAY kriging following IGNNK/GRIN, or LamaH-CE following Kirschstein & Sun (2024)?

2.The largest graph used has only 608 nodes. Could you test on a substantially larger graph, e.g. LargeST (~8.6k nodes), and report training time, memory, the number of GaBP iterations needed to converge, and accuracy against the baselines?

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

The authors proposed Gaussian Graphical Neural Networks (GGNNs), a novel GNN architecture that incorporates Gaussian graphical-model inference into the message-passing framework, specifically targetting the problem of estimating on a fixed spatial graph with sparse, noisy, or incomplete observations of node features. In this way, authors aim to leverage the representative power of GNNs while also utilizing the inductive bias of Gaussian graphical models to capture persistent dependencies between nodes across observations. The inference for the GMM module is performed using Gausian Belief Propagation (GaBP) on a latent space defined by a learned evidence matrix and a learned sparse precision matrix. The authors present three ways to parameterize the precision matrix (pairwise normal, diagonally dominant, and normalized Laplacian) based on walk-summability condition, ensuring the convergence of GaBP. All 3 parameterization methods are based on the fixed graph structure, and include node-specific and edge-specific learnable parameters, shared across all observations. The model is trained end-to-end using implicit differentiation, which allows efficient backpropagation through the GaBP iterations inference steps. The proposed method is evaluated on two tasks: daily temperature imputation and streamflow forecasting, showing competitive performance against several strong baselines.

Critical strengths
  • Interesting way to exploit the graph structure and combine with GGMs, with well-motivated and intuitive parameterizations of the precision matrix.
  • Experiment design is clear and consistent across models, and is able to highlight the benefits of GGNNs. Experiments also include a setting that strong baselines such as Transformer and GraphGPS outperform the proposed method.
  • Theorem 1 and corollary 1 show that walk-summability does not force a smoothing bias at any depth, which is a common problem for GNNs with multiple layers.
Critical weaknesses
  • The predictive benefits is outweighed by the computational cost, for example, on USCRN experiment, with positional encoding, GraphGPS consistently on-par or outperform GGNN with 4.4x faster training time, together with the number of learnable parameters scale number of nodes and edges of the graph limit the practical advantage of GGNNs.
  • Training protocol concern: On USCRN with or without coordinates input, GCN, IGNN and GPR-GNN do worse than a Transformer that does not use graph, and worse than the GGM in observation space, which seem to show under-training problems for these baselines, given a fixed learning rate, hidden size, and number of layers set by matching the number of parameters.
  • Missing a baseline able to learn long-range dependecies with learnable edges, which provide a clearer evidence to distinguish the benefit of learning edge weights with structured inference.
  • Code is not released.
Questions for authors
  • I would like to see a more thorough hyperparameter search for these baselines to ensure they are not under-trained, so that the claim is not confounded by the training protocol.
  • How was the standalone GGM baseline fitted on the fixed graph? How many layers for GCN-E? How many iterations for IGNN solver, how many hops for GPR-GNN?
  • The fixed 125 message-passing iteration should be justified.
  • Insights on when different parameterizations of the precision matrix better should be discussed, current results do not show a clear winner across tasks.
Reviewer_XcnEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Gaussian Graphical Neural Networks GGNNs for the setting where many observations share one fixed graph. The authors propose three parameterizations of the learned sparse precision matrix: pairwise-normalizable, diagonally dominant, and scaled Laplacian. Each guarantee walk-summability by construction, so Gaussian belief propagation (GaBP) converges. Training uses implicit differentiation that reuses the GaBP solver. Experiments cover USCRN temperature imputation and LamaH-CE streamflow forecasting.

Critical strengths

S1. Clear motivation. The case of fixed topology and repeated observations is common in environmental monitoring, and the paper ties it clearly to classical GMRF practice. S2. Constructive parameterizations. All three guarantee GaBP convergence with no projection or penalty.

Critical weaknesses

W1. Limited novelty. J is fixed across observations and h is linear, so each GGM sub-layer reduces to a learned linear graph filter. This is essentially the graph-signal-denoising view of GNNs with edge-specific weights. W2. Missing values are zero-filled into h rather than conditioned on, and because J does not depend on the observation, the model cannot raise precision at observed nodes, which is the very evidence-reconciliation mechanism that motivates the paper. W3. Baselines are weak. On USCRN, GCN, IGNN and GPR-GNN are nearly identical, around 2.5 times worse than a standalone GGM. GCN also barely improves with coordinates while the Transformer drops to 1.36, which suggests a pipeline or tuning problem, since all models share one hyperparameter setting. W4. The empirical advantage is modest and selectively reported. With positional encodings, Transformer wins RMSE at 12–48h on LamaH. Without positional encodings, GCN-E beats every GGNN variant on MAE at 72h and 96h. GGNN is also about 5.5 times slower than GCN.

Questions for authors

Q1. With receptive field matched, what remains of the gain? Q2. Can the marginal variances give calibrated uncertainty? That would clearly separate GGNN from deterministic implicit GNNs. Q3. Do the learned couplings reflect physical structure? Q4. Why is J independent of the observation?

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

This paper introduces Gaussian Graphical Neural Networks (GGNNs) for systems observed repeatedly over a fixed graph. Each block embeds a Gaussian graphical model in a residual GNN, learning a sparse precision matrix over the fixed edges and performing inference with Gaussian belief propagation. The authors propose three walk-summable precision parameterizations that guarantee convergent inference, show that this constraint imposes no inherent smoothing bias, and train the model via implicit differentiation. They evaluate GGNNs on spatial temperature imputation over 131 USCRN stations and streamflow forecasting over a 608-gauge river network from LamaH-CE. Without positional encodings, GGNNs generally outperform a GCN with learnable edge weights, reducing random-masking MAE from 1.38 to 1.32. With positional encodings, they remain competitive with GraphGPS and the Transformer.

Critical strengths
  • The problem setting is well motivated, and the method is designed around it. GGNN targets systems in which repeated observations share a fixed graph and learns persistent, edge-specific dependencies as a sparse precision matrix inside a GNN. The three parameterizations guarantee walk-summability by construction, so belief propagation converges by design, and the spectral analysis shows that this constraint does not force a smoothing bias. The implicit-differentiation scheme reuses the forward solver and reduces the memory complexity of training from O(T(E+N)d)O(T(E+N)d) to O((E+N)d)O((E+N)d), and GGNN’s peak GPU memory remains comparable to the baselines.

  • The experimental design isolates the effect of the graph operation. All methods share the same residual architecture and a budget of about 100k parameters, and results are averaged over 10 runs on a loopy spatial graph and a river tree, with and without positional encodings. GGNN consistently improves over a GCN with learnable edge weights on USCRN, including regional extrapolation (4.54 vs. 3.24 MAE when predicting the East from the West without positional encodings). The iteration ablation further shows that the amount of inference matters, as 24-hour MAE rises from 4.07 with exact tree inference to 4.59 with two iterations.

Critical weaknesses
  • The baselines omit the most closely related methods, and some appear undertuned. The introduction itself discusses spatiotemporal models that learn persistent adjacency, such as Graph WaveNet and AGCRN, and deep Gaussian Markov random fields, which are the closest prior work, yet none of them is evaluated. GAT, which learns edge-dependent weights through attention, is also absent. Meanwhile, GCN, IGNN, and GPR-GNN reach an MAE above 8 under random masking without positional encodings, worse than both the graph-agnostic Transformer and the linear standalone GGM. This makes their large gap to GCN-E hard to interpret and weakens the claim that learning edge-specific relationships matters more than the propagation mechanism.

  • The benefit of graphical inference is not consistent across settings. At the 72- and 96-hour horizons without positional encodings, GCN-E obtains a lower mean MAE than every GGNN variant with much smaller variance across runs, although the main text reports results only up to 48 hours. The iteration ablation is also non-monotonic at these horizons, where truncating inference to two iterations outperforms exact tree inference. With positional encodings, GraphGPS outperforms GGNN under both regional masks (e.g., 2.98 vs. 3.24 MAE). It therefore remains unclear when structured inference provides a reliable advantage, particularly at long horizons and when global attention is available.

Questions for authors

I have the following questions and suggestions for the authors.

  • Results for the closest related methods, such as Graph WaveNet and AGCRN for forecasting and deep Gaussian Markov random fields for spatial imputation, along with GAT, are needed to position GGNN against prior work that also learns persistent or edge-specific structure. It would also help to clarify whether hyperparameters were tuned for each baseline, as all models appear to share a single configuration. In particular, GCN, IGNN, and GPR-GNN perform worse than the linear standalone GGM under random masking without positional encodings (8.08 to 8.13 vs. 3.28 MAE), and the gap persists with positional encodings (7.36 for GCN vs. 1.36 for the Transformer), which suggests that it is not explained by the lack of node identity alone.

  • On USCRN, GGNN's advantage largely disappears once station coordinates are provided. With positional encodings, the Transformer improves from 5.38 to 1.36 MAE under random masking and GraphGPS outperforms GGNN under both regional masks, while GGNN barely benefits (1.32 to 1.31). One possible explanation is that GGNN's node- and edge-specific precision parameters already act as node identifiers, which the other baselines without positional encodings, apart from GCN-E, lack. Could the authors report baselines equipped with learnable node embeddings to separate the effect of node identity from that of graphical inference? Results combining GGNN with global attention would also clarify whether the two are complementary.

  • The LamaH-CE river network is directed from upstream to downstream, but its edges are symmetrized, presumably because the precision matrix must be symmetric. Since streamflow propagates downstream with a delay, this may discard information central to the forecasting task. Could the authors clarify whether the baselines also used the symmetrized graph and report baselines that use directed edges to quantify the cost of this constraint? A discussion of how GGNN could be extended to directed dependencies would also be helpful.
  • The evaluated graphs are small (131 and 608 nodes), yet GGNN is already more than four times slower per batch than GCN and GraphGPS, and USCRN requires a fixed budget of 125 iterations. Results on larger graphs, together with how the number of iterations needed for convergence scales with graph size and diameter, are needed to assess whether the method remains practical beyond these benchmarks.

  • Results at the 72- and 96-hour horizons should be reported in the main text. Without positional encodings, GCN-E achieves a lower mean MAE than every GGNN variant at both horizons (e.g., 8.52 vs. 8.69 at 96 hours), and GGNN shows much larger variance across runs (±0.82 for GGNN-L vs. ±0.06 for GCN-E at 96 hours). Could the authors explain this instability and why truncating inference to two iterations outperforms exact tree inference at these horizons (e.g., 8.58 vs. 8.94 at 96 hours)?

  • The reproducibility statement lists the code and dataset links only as placeholders (XXX.github.com and XXX.datasets.com). Could the authors provide anonymized code and the processed datasets, including the constructed USCRN graph, through the supplementary material or an anonymous repository?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.