acceptodds
Withdrawn from ICLR 2027

What Mean Aggregation Retains in GNNs: Degree, Precision, and Learned Readouts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mean aggregation in GNNs is non-injective, but this does not imply that it discards all information about neighborhood size. We characterize when degree remains recoverable from the aggregate and ask whether learned GNNs exploit that information. For categorical messages, the exact mean consists of empirical proportions (counts divided by degree), whose denominators constrain, and can sometimes identify, neighborhood size. Rounding or noise can remove this degree signal while preserving information about neighborhood composition. We use Fisher-information analysis to show how degree information depends on the message distribution: for Gaussian messages, the retained fraction vanishes at high degree even as information about the message mean remains. A broad class of regular readouts recovers asymptotically none of the degree information retained by exact categorical means; fitted neural readouts exhibit the same gap, and reducing numerical precision barely changes trained-GNN predictions. Supplying degree explicitly improves log loss and robustness under a low-degree shift but provides little benefit at high degree. Together, these results show that mean aggregation can retain degree information without learned GNNs making effective use of it.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
raph neural networks, mean aggregation, neighborhood size, degree information, information retention, Fisher information, readout functions, robustness, distribution shift

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

The paper studies graph neural networks with mean aggregation. Authors ask the question when, and to what extent, mean aggregation can be used to recover the information about the degree of a node, and how important is this information for a GNN. To study this problem, authors model the number of neighbours with a Poisson distribution. Then, they use Fisher information to determine how the expected value of the degree can be recovered from the mean aggregation of neighbours. In other words, authors study how well, on average, the degree of a node can be predicted from the mean aggregation of neighbours which are random. The paper contains both theoretical and experimental analysis.

Critical strengths

Mean GNNs, although very simple and very popular models are still not understood in full regarding their expressiveness and computational capabilities. A study of what their computations allow for, and their practical performance is both interesting and important.

The importance of the degree information – which is central structural information in graph data – is important to understand, and the paper provides another step in this direction.

The analysis of GNNs with Fisher information seems to be a novel approach, and may lead to interesting results. The direction is potentially promising and may open an interesting research direction.

Critical weaknesses

The way the paper is written, does not satisfy basic research paper standards. It is very unclear, imprecise, does not explain the setting and the meaning of the results appropriately; to the point that it is hard to judge the quality of the obtained results. The paper lacks appropriate explanation and definition of used concepts and is currently in the form which disallowed me to obtain its proper understanding.

The paper uses a number of undefined notions. For example “neighbourhood-category proportions” and “continuous-message centres” in 055 are never explained before. Similarly in the description of contributions it is not clear what are “exact proportions”, what is “composition”, what does it mean that “realized degrees remain ambiguous”, what is meant by “rounding or additive noise” as it is not explained where this noise is added, what is meant by a readout (which is not as in standard GNNs) and in particular what are “regular readouts” or ”fitted neural readouts”.

Statement “At finite expected degree” remains unclear for me. Also, “location information” is not explained and what does it mean that “it survives” is unclear. Furthermore “degree-dependent grid of possible proportions” is not explained. Similarly “complete neighbourhood” in 149 probably should be “complete observation”, but I am not sure. “Output H” and “aggregation channels” are not introduced in the paper before the usage of these terms, and so is “parameter-independent channel with regular support” which occurs in Proposition 1. Similarly “fixed probe feature map” in Proposition 2 is not defined.

Experimental section is also unclear for me, namely the whole setting seems to be underspecified in the main body of the paper, where it should be explained. For example it is not clear on what graphs GNNs are trained and on what task they are trained. Also it is not clear for me what is the relation of the experiments to the theory. The results presented in “Explicit degree under degree shift.” and “Sensitivity to aggregate precision and noise.“ do not have a discussion that would link them to theoretical results, which remains unclear for me.

Questions for authors

I wonder if and how the experimental results in “Explicit degree under degree shift.” and “Sensitivity to aggregate precision and noise.“ are related to the theoretical results in the paper.

Reviewer_bivAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work studies whether mean aggregation preserves information about the degree of the neighborhood, and whether this can be extracted by the model. They analyze how much is preserved for categorical and gaussian messages; then, they study how much is lost due to reducing precision; finally, they do tests on synthetic graphs with the degree as input, where there is a distribution shift between training and test.

Critical strengths

The paper is complete in terms of ablations and covers many settings.

Critical weaknesses

The conclusions are a bit underwhelming: yes, sometimes the degree gets lost with mean aggregation, and yes, including it can help, in the case that the task would require it. I would try to extend the conclusions to something more readable, instead of a checklist. For instance, I would like to see the conclusions impacting further work. I see a problem with the work not stating its own relevance or context.

Also, I don't see it making a lot of difference for many real world tasks. For instance, in recent work (Rubio-Madrigal et al., ICML 2026) it's shown that repeated mean aggregation already performs well for many node classification tasks, and they include an analysis of what is potentially lost. The empirical conclusion here is very limited, but carries a lot of pull on the storyline. For instance, the node classification experiments are just two layer narrow GNNs on SBMs, I assume, without tuning. And the information recovery is not done on those GNNs, but on the aggregation.

Questions for authors

In which downstream tasks does the degree missing from the mean limit prediction in the end? Where could you see, e.g., precision making a difference for prediction?

How were the optimization details chosen?

Reviewer_1pn2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper applies Fisher information to answer the question of how much information the mean aggregation retains in GNNs in a Poisson-distribution neighborhood model. The paper's findings include: i) Exact categorical proportions asymptotically retain all local degree and composition information, ii) Additive noise or rounding erases degree information but keeps composition, iii) Regular readouts recovers an asymptotically vanishing fraction of the information retained by exact categorical means, and iv) Gaussian means retain a vanishing degree fraction.

Critical strengths

The primary strength of this paper is its mathematical competence, with several theorems, corollaries, and propositions. However, I don't think that the findings are significant enough because the issue addressed in the paper is a simple observation. The Fisher information theory does not help much.

Critical weaknesses

Several claims in the paper, such as “none measures how much degree information a normalized aggregate retains” or “no prior work identifies when a perturbation erases that arithmetic signal”, could be incorrect. The experiment does not prove the finding. According to Thms 6–7, the mean retains no degree information. In the experiment, adding degree helps because it supplies information the aggregator discards. But it is the opposite of what was claimed before “retained but unused.”

Questions for authors

Is it possible to train a GNN on categorical features with explicit low-precision training where the arithmetic channel matters?

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

Related papers

Discussion (0)

Sign in to comment.