acceptodds
Under review as a conference paper at ICLR 2027

Seeing More, Distinguishing Less? Learning from Graph Propagation States and Changes

Abstract

Decoupled graph neural networks precompute multi-hop features through graph diffusion, enabling information from different propagation scales to be used without increasing neural network depth. However, existing cross-hop fusion methods often assign a single weight to each propagation step and share it across feature dimensions, ignoring that different features could benefit from different propagation scales. Moreover, using more propagation scales does not necessarily improve node distinguishability, as combining multi-hop states can weaken useful differences across propagation steps. To this end, we propose Feature-wise Fusion across Hops for Adaptive Graph Learning (FOCAL), a graph learning framework that models Propagation States and Changes through feature-wise cross-hop fusion. In the method, Propagation States describe feature values after diffusion, while Propagation Changes capture their differences between consecutive diffusion steps. FOCAL learns independent weights for each feature in the State and Change branches, allowing different features to use different combinations of propagation operators and diffusion steps. Node-level gates further adjust the contributions of the encoded State and Change representations for each node. Experiments on eight node classification datasets show that FOCAL achieves the highest accuracy on seven datasets. Ablation studies and weight analysis further demonstrate the contributions of Propagation Changes, feature-wise fusion, separate State and Change weights, and Node gates.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural network, graph diffusion, graph learning

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

The paper argues that in decoupled GNNs the usual cross-hop fusion of precomputed propagation states H^(k) = P^k X is mis-designed in two ways: it assigns one weight per propagation step shared across all feature dimensions, and it optimises only the states, so that combining many hops can actually reduce node distinguishability. FOCAL therefore fuses two families of signals separately — Propagation States and Propagation Changes Δ^(k) = H^(k−1) − H^(k) = (I_n − P_t)H^(k−1) — using two independent per-feature weight matrices Θ_u, Θ_v ∈ R^{d×2K} (softmax-normalised over the two operators {self-loop, non-self-loop} and K hops, with a 2^{-k} prior), then encodes the input, state and change representations with three separate encoders and combines the latter two through per-node gates. Experiments on eight node-classification datasets with a single shared configuration report the best mean accuracy on seven, with ablations over the fusion and gating design and an analysis of the learned hop weights.

Critical strengths
  • Separating Propagation States from Propagation Changes, and showing that equal-weight state mixing merges nodes that change mixing separates, is a crisp and well-motivated idea.
  • Per-feature weights over operator×hop generalise the GPR-GNN/FSGNN scalars at O(Knd) extra cost, with a sensible 2^{-k} geometric prior.
Critical weaknesses
  1. The central distinguishability claim is neither formalised nor measured, resting on a hand-built four-node example with no separation or Dirichlet-energy metric anywhere in the experiments.
  2. The headline gains are within one standard deviation of the strongest baselines on five datasets, while the large Squirrel/Cornell/Texas jumps rest on non-standard random splits and K = 5 is selected on test accuracy.
  3. The two distinguishing designs add little (Global-Fusion within 0.3–1.6 points and State-State within ~1.3 points on five of eight datasets), so extra capacity is not separated from the change signal.
  4. Rigour gaps: baseline hyperparameters come from other papers' splits, the strong heterophily competitors (H2GCN, GloGNN, OrderedGNN, GBK-GNN) are missing exactly where the gains are claimed, and there is no parameter/FLOP/runtime table, code or reproducibility statement.
  5. The learned-weight analysis shows heterogeneity but never establishes causality — no effect size, no correlation with per-dataset gain, and no interventional check that the scale preferences cause the improvement.
Questions for authors
  1. The central distinguishability claim is neither formalised nor measured, resting on a hand-built four-node example with no separation or Dirichlet-energy metric anywhere in the experiments.
  2. The headline gains are within one standard deviation of the strongest baselines on five datasets, while the large Squirrel/Cornell/Texas jumps rest on non-standard random splits and K = 5 is selected on test accuracy.
  3. The two distinguishing designs add little (Global-Fusion within 0.3–1.6 points and State-State within ~1.3 points on five of eight datasets), so extra capacity is not separated from the change signal.
  4. Rigour gaps: baseline hyperparameters come from other papers' splits, the strong heterophily competitors (H2GCN, GloGNN, OrderedGNN, GBK-GNN) are missing exactly where the gains are claimed, and there is no parameter/FLOP/runtime table, code or reproducibility statement.
  5. The learned-weight analysis shows heterogeneity but never establishes causality — no effect size, no correlation with per-dataset gain, and no interventional check that the scale preferences cause the improvement.
Reviewer_4LZROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FOCAL, a decoupled graph neural network for transductive node classification.

Critical strengths

The architecture offers a possible way to combine feature-dependent propagation with node-dependent weighting.

Critical weaknesses
  1. Writing. 1.1. In Figure 2, why does the graph of P_s have duplicated nodes? What does it mean? 1.2. In Figure 2, what does graph diffusion mean?

  2. Methodology. 2.1. "For t \in {n, s}", what do "n" and "s" mean? If "n and s denote nonself-loop and self-loop propagation", then I don't understand how graph diffusion can be unified in Eq. 1. 2.2. "The representation used for prediction is Z_i=..." I don't understand this part. 2.3. In Eq. 6, what are W^g2 and b^g2?

  3. Experiments. 3.1. Many baseline performances are different from the numbers reported in their papers. Such as, ACM and FSGNN achieved much higher results than declared in their papers. 3.2. I checked the authors' uploaded code; there are no baseline method implementations.

Questions for authors

See Critical Weaknesses

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

This work augments "multi-hop fusion for decoupled GNNs" with two components: a dedicated hop-weight set for each individual feature, plus an auxiliary branch capturing the difference between consecutive hops.

Critical strengths

Sound engineering design, clear writing, and comprehensive ablation studies. It yields 0.5–1.5% accuracy gains on homophilic graphs (Cora, CiteSeer, PubMed, WikiCS), and substantial improvements on heterophilic graphs.

Critical weaknesses

The opening motivating example does not adequately support the paper’s rationale. Nodes A and B are averaged into identical values only when State uses equal weights; as soon as weights become unequal (and FOCAL itself learns non-uniform weights), A and B are immediately distinguishable. Therefore, the example merely demonstrates "equal weighting should be avoided", rather than justifying the need for a variation branch.

For the four heterophilic graphs, the code uses PyG’s directed edge lists, yet the paper defines the operator for undirected graphs and reports undirected deduplicated edge counts in tables. Furthermore, the repository lacks fixed random seeds, per-split results, and baseline implementations. When I reran Cornell and Texas using undirected edges, the accuracy dropped by 4.8 and 2.9 points respectively. This issue requires further detailed clarification from the authors.

Questions for authors

Two weaknesses requires further detailed clarification from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.