acceptodds
Under review as a conference paper at ICLR 2027

Beyond Distance: Reliability-Gated Fisher–Rao Routing for Heterophilic Graphs

Abstract

Heterophilic graphs connect nodes with different labels, yet nodes of the same class can retain characteristic neighborhood label distributions (NLDs). Existing methods primarily treat NLDs as Euclidean statistics, auxiliary targets, or pairwise similarity signals. We instead place NLDs on the probability simplex and distinguish the radial magnitude of a Fisher–Rao displacement from its tangent direction. This motivates FR-NLD, a reliability-gated geometric network in which every neighbor acts as a viewer of the target node and candidate class prototype. The model compares the target and prototype in the Fisher tangent space anchored at the neighbor, using angular evidence only when both displacements are sufficiently reliable and otherwise falling back to radial evidence. Degree-aware NLD estimation regularizes finite neighborhoods, while residual routing and belief refresh support repeated geometric inference. Experiments on five modern heterophily benchmarks using all ten public splits show that FR-NLD achieves the highest Minesweeper score among the recent published comparators in our evaluation and remains competitive on Roman-empire and Tolokers. On Minesweeper, it outperforms reproduced Polynormer while using only 21% of its parameters and peak GPU memory. Controlled checkpoint interventions further show that predictions depend on the neighbor-specific tangent reference and angular evidence, supporting the proposed geometric mechanism.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
heterophilic graphs, graph neural networks, information geometry, Fisher–Rao geometry, neighborhood label distributions, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_ygPhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FR-NLD addresses node classification on heterophilic graphs by placing neighborhood label distributions, estimated from soft predictions, on the probability simplex. For a neighbor, target node, and candidate class, it compares directions toward the target and class prototype in the Fisher–Rao tangent space anchored at the neighbor. Radial distances gate angular evidence, producing signed, class-conditioned messages. Dynamic configurations additionally use degree-aware shrinkage, residual inference, and periodic belief refresh; Amazon-ratings uses a different feature–geometry fusion path.

The authors evaluate on ten public splits of five modern heterophily benchmarks, distinguishing a single model, a three-member Plus variant, and a Pro variant incorporating auxiliary GCN predictions. The single model reaches 99.116% ROC-AUC on Minesweeper; Pro has the highest mean on three datasets among the ranked comparators in Table 1. The paper also reports same-hardware resource comparisons, compact-model interventions, and appendix analyses of priors, refresh, and depth.

Critical strengths
  1. The router has a clear technical motivation and some supporting mechanistic evidence. Equations (3)–(4) on page 5 compare class directions in a neighbor's tangent space and fall back to radial experts for short displacements, yielding interpretable class-conditioned messages. Appendix D.3 reports a 0.161 percentage-point improvement from the combined prior on Roman-empire, with a paired 95% interval of [0.015, 0.306]. An earlier matched experiment reports a 0.607-point benefit from dynamic refresh, with an interval of [0.388, 0.827]. These support the prior and iterative-inference components, although they do not establish the necessity of angle or gating themselves.

  2. The single-model Minesweeper result and its particular resource savings are meaningful. Table 1 reports 99.116±0.170. Against the authors' same-hardware Polynormer reproduction in Table 2, which obtains 97.281, the gain is 1.835 percentage points, with 2.20M versus 10.54M parameters and 0.97GB versus 4.15GB peak memory. This is strong evidence that does not depend on Pro ensembling. The use of five modern datasets and ten public splits, together with disclosure of negative ablations and resource disadvantages on Amazon-ratings, also improves transparency.

Critical weaknesses
  1. The independent contributions of direction and reliability gating remain insufficiently quantified. The geometry-off intervention in Table 3 removes a substantial routing module, including class-conditioned messages and prototype-related inputs; the full performance gap cannot be assigned to angle alone. After retraining without angle, the gaps are +0.236, +0.118, +0.012, and +0.313 percentage points on Roman-empire, Minesweeper, Tolokers, and Questions, respectively, but −0.261 on Amazon-ratings. The authors candidly acknowledge these mixed outcomes. What is missing is uncertainty for these particular contrasts and validation in at least one representative main configuration, rather than a wholesale repetition of already completed ablations. Zeroing angles or shuffling viewers in a frozen model shows dependence on those inputs, but changes the input distribution and cannot by itself establish that a retrained alternative router is worse. The prior/depth analyses also do not directly isolate the three-expert reliability gate.

    There is a further interpretive boundary. The simplex tangent space is one-dimensional for the three binary datasets. For nondegenerate displacements, the Fisher–Rao tangent cosine equals

    sign⁡[(ru,1−rv,1)(μk,1−rv,1)]∈{−1,+1}.\operatorname{sign}\big[(r_{u,1}-r_{v,1})(\mu_{k,1}-r_{v,1})\big]\in\{-1,+1\}.

    This remains useful directional information, but it is exactly the same as this ordering test; treating the equivalent expression as an independent competing baseline would be uninformative. The binary results can support the complete router, but cannot distinguish this angle from a one-dimensional Euclidean angle. A discriminating test should instead use a multiclass task, preserving prototypes, message channels, and training budget while comparing a radial router or a reasonable Euclidean geometry, and separately examining the gate. Likewise, the square-root map with spherical geodesic geometry preserved is an isometric reparameterization of Fisher–Rao, not an independent alternative.

  2. System rankings, evaluation protocols, and efficiency claims need clearer separation. Pro benefits from multiple FR-NLD members and auxiliary GCN predictions; Appendix Table 4 explicitly supplies mixture weights for three auxiliary GCNs. In Table 1, the single FR-NLD obtains 91.677 versus Polynormer-r's 92.55 on Roman-empire and 78.405 versus 78.92 on Questions. Table 2 uses a different, same-hardware Polynormer reproduction, obtaining 78.299 on Questions; conclusions from these two tables should not be conflated. Table 1 also mixes ten-split results with source-specific protocols. Pro's highest means do not automatically establish either the superiority of the core router or statistical significance.

    On Amazon-ratings, Table 2 reports 162% of the comparator's parameters and 199% of its memory, with lower performance. The Minesweeper efficiency advantage therefore does not generalize to every configuration. Appendix E.1 defers the complete search candidates and validation rankings to release, limiting scrutiny of search budgets. Useful additions would be the auxiliary predictor's own performance, one comparably budgeted ensemble control, and complete deployment costs. Ensembling and validation-based tuning are not themselves methodological errors.

Questions for authors
  1. Primary; could affect the recommendation: Please provide per-split results or paired intervals for the retrained angle contrasts in Table 3. In one representative multiclass main configuration, can you compare radial/Euclidean routing while holding class prototypes, message channels, and budget fixed? What does the three-expert reliability gate contribute relative to a control without that distance-dependent gating? A separate evaluation of the mathematically equivalent binary sign implementation is not requested.
  2. Primary; could affect the recommendation: What is Pro's incremental benefit over the auxiliary GCN predictor alone and one comparably budgeted baseline ensemble? Please distinguish the source results in Table 1 from the same-hardware reproduction in Table 2, and report actual Plus/Pro inference costs and paired differences for the principal comparisons.
  3. Reproducibility clarification: Please supply the A28 specification, stopping criteria, number of candidate configurations, and validation-selection record. The paper mentions anonymous supplementary utilities, but they were not included in the PDF materials supplied for this review; if they contain these details, please identify where.
  4. Interpretive clarification: As displacements shrink, direction may persist while fixed positive thresholds suppress angular weight. What are the observed displacement-to-threshold ratios and expert masses during training? This would clarify the useful operating regime. I do not interpret the local motivation as an already-established universal anti-oversmoothing theorem.

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

The authors propose FR-NLD, a graph neural network for heterophilic graphs that uses the Fisher-Rao divergence between neighborhood label distributions to capture information about adjacent vertices that Euclidean distance cannot. The authors evaluate their method on a standard benchmark suite against comprehensive baselines, reporting state-of-the-art performance on heterophilic graph node label prediction performance.

Critical strengths

The paper's main strength is in its empirical results for FR-NLD being competitive with the state-of-the-art in heterophilic graph prediction. The authors present a comprehensive list of baselines across a list of standard benchmarks, and show that FR-NLD performs comparably to a state-of-the-art baseline with fewer parameters and VRAM requirements.

Critical weaknesses

The paper's primary issue is its lack of clarity. First, the authors assume familiarity with differential geometry in their presentation, which is not suitable for the general ML audience at ICLR; this could be fixed by presenting the work differently, as the subject matter itself is suitable for ICLR. The notation is overly dense, so understanding the paper is difficult even with careful reading. Terms are used before they are introduced properly (see "routing blocks" and "index routing depth" in the first paragraph of section 3, and "feature-prototype scores" in the last paragraph of section 3.2 for a few non-exhaustive examples). Figure 1 is too small, and even when zoomed in does not help my understanding of the method. Entire methodological descriptions are omitted (for example, what is being distilled from a GCN? Distillation is mentioned a couple of times in the experiments section, but nowhere else in the paper body). There is no motivation given for many of the design decisions made in the paper; rather, terms are introduced and used in equations without justifying to the reader why their inclusion in the final FR-NLD architecture is necessary. These presentation issues are significant enough that I feel I cannot provide an accurate evaluation of the technical correctness of this paper. Therefore, I am quite confident that the presentation issues of FR-NLD significantly reduce its utility to the scientific community, regardless of any potential contributions.

A second critical weakness is in the evaluation of FR-NLD. FR-NLD-plus is an ensemble of three FR-NLD models trained on seemingly the same data but with different random seeds, but some datasets use different selection criteria based on validation data than other datasets to select how the ensemble works---e.g., the Questions dataset seems to train 4 potential ensemble models and then choose the top 3 from validation data, whereas the other datasets directly train 3 ensembles. Other datasets differ on how the ensemble weights are selected Therefore, FR-NLD-plus results compared to baselines in Table 1 should be considered invalid due to a per-dataset tuning not available to the baselines.

FR-NLD-pro predictions are simply a mixture of FR-NLD-plus predictions and GCN predictions as \rho * (FR-NLD-plus) + (1 - \rho) * (GCN), where \rho is a mixing weight determined from validation data, and therefore inherits the dataset inconsistency that FR-NLD-plus has. At minimum the FR-NLD-plus and -pro experiments are invalid to include as fair comparisons against the baselines. With -plus and -pro omitted, FR-NLD results as presented provide no compelling advancement over the state-of-the-art.

Questions for authors

Clarification: FR-NLD -plus and -pro variants are not specific to FR-NLD. Any of the baselines could be trained as ensembles and then mixed with a GCN to create their own -plus and -pro variants. In a fair comparison where you give your baselines these ensemble and mixing strategies with the same tuning access to validation data, does FR-NLD still outperform baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.