acceptodds
Withdrawn from ICLR 2027

From Euclidean to Riemannian: Boundary-Aware Geodesic Oversampling for Class-Imbalanced Graphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Class imbalance is a widespread problem in graph-structured data. A promising solution is to oversample minority samples to balance the class distribution. However, existing methods mainly synthesize samples via Euclidean linear interpolation, which deviates from the intrinsic geometry of graph representations. Moreover, even the interpolation between intra-category samples passes through regions dominated by other categories, while improperly synthesized edges further distort the local topology. To address these limitations, we propose a oundary-ware iemannian eodesic versampling method () for class-imbalanced graph learning. Specifically, we first employ a Riemannian graph encoder to map nodes onto a geometry tailored to the hierarchical structures of graphs and construct Fréchet prototypes to characterize class-wise distributions. Then, we sample reliable minority seed and partner nodes, and synthesize candidate nodes along their intrinsic geodesics. Furthermore, we develop a dual screening method that evaluates both endpoint margins and intermediate points along the generation path, thereby reducing the risk of crossing category boundaries. Finally, we propose a geometry-aware topology generator. Experiments on six real-world datasets demonstrate the effectiveness of the proposed method.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Representation Learning, Class Imbalance, Riemannian Manifold, Node Classification

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 3 reviews, above 27% of the others.
Reviewer_zd5ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BARGO, a graph oversampling method that generates minority nodes along geodesics in a Riemannian latent space. It combines reliable seed selection, endpoint and path screening, and a topology generator for synthetic nodes. Experiments on six artificially imbalanced benchmarks report consistent improvements over loss-reweighting and graph-oversampling baselines.

Critical strengths
  1. BARGO considers representation synthesis, boundary crossing, and synthetic-node topology together rather than treating generated nodes as independent feature vectors.

  2. The evaluation includes loss-modifying, oversampling, generative, and topology-augmentation methods across six datasets and reports means and standard deviations over ten runs.

  3. BARGO achieves the best reported mean results across both long-tailed and step-imbalance settings, and the component ablations indicate that reliable sampling, screening, and topology coordination are useful.

Critical weaknesses
  1. The central Riemannian contribution is not isolated. The paper does not compare BARGO with an otherwise identical Euclidean version that retains the same seed selection, endpoint/path screening, oversampling amount, and topology generator. The ablations remove screening or topology modules, but none replaces the Lorentz encoder and geodesic interpolation with a Euclidean encoder and linear interpolation. Therefore, the reported gains cannot be attributed to Riemannian geometry, despite this being the main claim and title of the paper.

  2. The experimental protocol is incomplete and internally inconsistent. The paper does not report the train/validation/test splits, labeled samples per class, hyperparameter selection procedure, manifold curvature, or most thresholds used in Equations (10)–(23). Section 4.1 also claims that Appendix A.1 contains experiments on three naturally imbalanced datasets, but the appendix provides no such experiments. In addition, the appendix refers to the proposed method as “GraphCALH” and leaves the CB Loss citation unresolved. These omissions prevent reliable reproduction and weaken the claimed generalization beyond manually imbalanced benchmarks.

Questions for authors
  1. How does BARGO compare with a matched Euclidean model that uses the same screening and topology modules but replaces geodesics with linear interpolation?

  2. What fraction of Euclidean and Riemannian interpolation paths violate the endpoint or path criteria, and how often does BARGO shorten or reject a candidate?

  3. Can the authors report endpoint-only and path-only screening ablations? The current joint removal does not establish the additional value of checking intermediate geodesic points.

  4. What are the exact data splits, manifold curvature, threshold values, and tuning protocol, and where are the promised experiments on naturally imbalanced datasets?

Reviewer_i36kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Boundary-Aware Riemannian Geodesic Oversampling (BARGO) for class-imbalanced node classification. The method employs a Riemannian graph encoder and Fréchet prototypes to represent class distributions, selects reliable minority samples, and generates synthetic nodes through geodesic interpolation. It further introduces endpoint-level and path-level screening to reduce cross-boundary interpolation and a geometry-aware topology generator to construct edges for synthetic nodes. Experiments on six benchmark datasets demonstrate improvements over existing methods.

Critical strengths

S1. The proposed framework jointly considers representation geometry, boundary reliability, and topology consistency, providing a systematic approach to minority-node augmentation.

S2. The endpoint-level and path-level screening strategy is an interesting design for reducing the risk of cross-boundary interpolation.

S3 The experimental results are generally positive across six datasets, and the ablation studies demonstrate the effectiveness of several proposed components.

Critical weaknesses

W1. Insufficient validation of Riemannian geometry. The paper lacks a direct comparison between Euclidean and Riemannian interpolation under identical sampling and screening strategies. Therefore, it is unclear whether the improvements primarily come from Riemannian geometry or other proposed modules.

W2. Potential selection bias in minority sampling. The reliability-based sampling strategy favors high-confidence minority nodes, potentially excluding boundary samples and reducing minority-class diversity. The impact of this design is not investigated. It remains unclear whether the generated nodes effectively improve minority-class discrimination or merely increase sample density in high-confidence regions. Additional comparisons with hard-sample generation methods, such as GraphSHA, and analyses of the distribution of synthetic nodes relative to decision boundaries would strengthen the paper.

W3. Insufficient experimental evaluation. The experiments are limited to six standard datasets with manually constructed class imbalance, without considering established benchmarks such as IGL-Bench. The dataset selection and experimental settings are insufficient to demonstrate the generality of BARGO. More comprehensive evaluations on diverse graph datasets and imbalance settings are needed.

W4. Incomplete related work and unclear novelty. The paper overlooks relevant work, such as Hyperbolic Geometric Graph Representation Learning for Hierarchy-imbalance Node Classification, which has already explored hyperbolic geometry for imbalanced graph learning. The authors should discuss this work and better clarify the novelty of BARGO relative to existing geometry-based approaches.

W5. Insufficient validation of topology generation. The use of local graph curvature and degree calibration is not sufficiently justified, and the paper lacks direct evaluations of the structural quality of generated edges.

Questions for authors

See the above

Reviewer_YEQ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles class-imbalanced node classification on graphs. It argues that existing oversampling methods (e.g., GraphSMOTE, GraphSHA) rely on Euclidean linear interpolation, which deviates from the intrinsic geometry of graph representations, and that even intra-class interpolation paths may traverse majority-dominated regions.

Critical strengths

The observation that endpoint reliability does not imply trajectory reliability is an under-explored failure mode of interpolation-based oversampling. The endpoint-level+path-level dual screening with adaptive shortening follows naturally from this diagnosis and, to my knowledge, is new in graph oversampling.

Critical weaknesses

Riemannian/hyperbolic GNN encoders, Fréchet means as class prototypes, and geodesic interpolation are all established components; the paper's contribution is their careful combination plus heuristic thresholding. Critically, prior work already combines hyperbolic geometry with imbalanced learning yet none are cited or compared. Since the paper's central claim is that geometry matters for imbalance, the absence of a geometry-aware imbalance baseline leaves the key claim unsubstantiated against its nearest competitor.

Section 4.1 states "additional experiments on three naturally imbalanced datasets are provided in Appendix A.1", but A.1 contains only dataset and baseline descriptions-the promised experiments are absent.

How the Fréchet prototype is computed (closed-form Lorentz midpoint vs. iterative Karcher flow) with as few as n_max/100 minority labels; whether manifold curvature is fixed or learned; the definition of "local graph curvature" κ_k in Eq. (22); and all thresholds (η_r, η_e, ξ_e, m^e_c, m^path_c, τ, δ±).

Questions for authors
  • How does BARGO compare to hyperbolic imbalanced-learning methods (e.g., HyperIMBA) under your protocols? If such baselines underperform, why?

  • Where are the naturally-imbalanced-dataset experiments promised in Section 4.1? Appendix A.1 even mentions ogbn-arXiv, but no results are reported anywhere.

  • How is the Fréchet prototype computed, and is it stable when a minority class has only a handful of labeled nodes (ρ=100)?

Comments on the paper

2 comments

Thank you for the response, which addresses most of my concerns. However, the response to Question 2 does not match the original question: it compares only different datasets, without separately reporting violation rates for Euclidean versus Riemannian paths or for endpoint versus path criteria, and without distinguishing path shortening from candidate rejection. Therefore, the current results do not establish that Riemannian paths are more reliable than Euclidean paths and are insufficient to raise my rating to weak acceptance.

Thank you for the additional response, which resolves many of my concerns. The new results answer my question, but they also raise a more fundamental concern: Riemannian paths fail the screening criteria far more often than Euclidean paths. This suggests that the gain may come from aggressive screening and shortening rather than from more reliable Riemannian interpolation. The comparison is also difficult to interpret because the two variants retain very different numbers of synthetic samples. For example, on Cora-LT, 8,366 Riemannian proposals are accepted, compared with 22,741 Euclidean proposals. Unless the final sample counts and class distributions are matched before training, the ablation changes both the interpolation geometry and the training data. The contribution of Riemannian geometry therefore remains unclear, so I maintain my original score.

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

Related papers

Discussion (0)

Sign in to comment.