acceptodds
Under review as a conference paper at ICLR 2027

CuBAS: A New Curvature-Based Adaptive Sampling Method for Supervised Classification

Abstract

The informativeness of a training set is often more important than its size. However, existing sampling strategies largely ignore the intrinsic geometry of labeled data. We introduce CuBAS (curvature-based adaptive sampling), a information-geometric framework for adaptive sample selection in supervised classification. The key insight of the new method is that a labeled dataset induces a statistical manifold whose local curvature provides a principled measure of sample informativeness. Specifically, we model label interactions through a -state Potts Markov random field defined on a -nearest-neighbor graph, and derive a closed-form curvature estimator from the observed Fisher information. Adaptive thresholding of the resulting curvature signal partitions the graph into smooth low-curvature regions – well represented by a small number of prototypes, and high-curvature regions concentrated around class interfaces – where samples are most informative for defining decision boundaries. While jointly sampling from both regimes, CuBAS constructs compact yet highly informative training subsets that preserve global representativeness as well as local discriminative structure. Extensive experiments on more than 60 benchmark datasets spanning tabular, image, and biological domains demonstrate consistent and statistically significant improvements over random and entropy-based sampling across a wide range of labeling budgets. Beyond its empirical effectiveness, CuBAS provides a computationally efficient and theoretically grounded connection between information geometry and adaptive data selection for supervised learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Curvature-based adaptive sampling, Information geometry, Fisher information, Active sample selection, Supervised classification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MpC9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CuBAS, a sample selection method for supervised classification. It constructs a k-nearest-neighbor graph over labeled data and models local label interactions using a Potts Markov random field. A ratio of node-wise Fisher-information-related quantities is interpreted as a curvature score. An adaptive threshold separates samples into low- and high-score groups, from which equal proportions are sampled to form a training subset. Experiments evaluate a k-NN classifier on 68 benchmark datasets across multiple training fractions. The paper reports substantial improvements over random and entropy-based sampling, attributing these gains to preserving boundary information and representativeness while reducing redundancy.

Critical strengths
  1. The motivation to balance decision-boundary information with broad data representativeness is reasonable. Identifying informative samples while avoiding redundant selections is an important problem, and local label interactions provide a potentially useful signal.

  2. The paper reports substantial improvements over random and entropy-based sampling across 68 datasets and multiple training budgets. The breadth of datasets and repeated evaluations provide a useful empirical starting point for investigating the proposed approach.

Critical weaknesses
  1. Unclear connection between the score and boundary informativeness. The paper does not establish why its score identifies boundary samples or redundant regions. For a binary neighborhood with class counts a,ba,b and observed label A, ignoring regularization gives ∣Si∣=eβ(a−b)|S_i|=e^{\beta(a-b)}. For positive β\beta, stronger agreement with the observed label increases this magnitude, challenging the interpretation that high ∣Si∣|S_i| indicates class interfaces while homogeneous regions have near-zero curvature.

  2. Insufficient justification of the curvature interpretation. Under regularity conditions, the expected squared score and expected negative log-likelihood Hessian are two expressions of the same Fisher information. Their observed counterparts can differ, but this does not establish them as first and second fundamental forms. The appendix derives these quantities algebraically without providing the geometric construction needed to identify their ratio as a shape operator.

  3. Unsubstantiated redundancy reduction and insufficient baselines. Algorithm 1 samples the same proportion from both strata, without explicitly discouraging selection of mutually similar samples. Consequently, the claimed redundancy reduction and the specific contribution of curvature remain unclear. CCS[1] already studies score-based stratification, while D² Pruning[2] and SES[3] explicitly incorporate diversity mechanisms. Comparisons only against random and entropy sampling cannot establish the advantage of the proposed score and sampling combination.

References

[1] Zheng et al. Coverage-centric Coreset Selection for High Pruning Rates. ICLR 2023.

[2] Maharana et al. D² Pruning: Message Passing for Balancing Diversity and Difficulty in Data Pruning. ICLR 2024.

[3] Xie et al. Structural-Entropy-Based Sample Selection for Efficient and Effective Learning. ICLR 2025.

Questions for authors
  1. Could you provide the geometric construction justifying the proposed shape operator and explain the binary-neighborhood example above? Please clarify whether normalization uses SiS_i or ∣Si∣|S_i|, and provide evidence connecting the resulting score to boundary proximity and redundancy.

  2. Can you disentangle the contributions of the score and sampling procedure? Please replace curvature with simpler neighborhood statistics under the same stratification and allocation rules. What mechanism or measurement supports redundancy reduction beyond preserving the proportions of the two strata?

  3. Could you compare against CCS[1] and representative diversity-aware methods, such as D² Pruning[2] or SES[3], on a manageable subset of representative datasets and budgets? A complete rerun on all datasets is unnecessary, but stronger comparisons under matched information access and feature representations are needed.

  4. Is a common independent test set held out before graph construction, parameter estimation, and sample selection? Algorithm 1 appears to use the full labeled dataset before assigning the remainder to testing. Please clarify this protocol and, if necessary, reevaluate without using test labels.

I would be willing to raise my score if the authors convincingly address Questions 1 and 2 with clear theoretical justification and supporting evidence.

References

[1] Zheng et al. Coverage-centric Coreset Selection for High Pruning Rates. ICLR 2023.

[2] Maharana et al. D² Pruning: Message Passing for Balancing Diversity and Difficulty in Data Pruning. ICLR 2024.

[3] Xie et al. Structural-Entropy-Based Sample Selection for Efficient and Effective Learning. ICLR 2025.

1 comment

Thank you for your thorough review and for the time and effort you put into your comments, questions, and suggested references. We appreciate your acknowledgment of the motivation behind balancing boundary information and representativeness, as well as the breadth of our empirical evaluation. We are working on the paper and will implement any pertinent improvements in the revision.

Reviewer_4NzSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CuBAS constructs a labeled k-NN graph, fits a Potts model, and derives a Fisher-information-based score to partition samples into two groups for training-subset selection. An adaptive threshold determines the partition.

Critical strengths

The scoring procedure uses graph-based label statistics without requiring an auxiliary predictive model.Experiments cover 68 datasets, 15 training fractions, and 100 repetitions per setting, providing substantial coverage across data regimes.

Critical weaknesses

(1) Algorithm 1 computes label-dependent scores on the entire dataset before returning unselected samples as the test set. Consequently, test labels influence training-subset selection. Evaluation should instead hold out an independent, common test set before graph construction and selection. (2) Section 3 treats expected squared scores and expected negative log-likelihood Hessians as distinct geometric quantities, although they are equivalent Fisher-information expressions under the stated regular conditional model. This identity is also given in the cited information-geometry literature. Replacing expectations with per-node observations does not establish that their ratio in Eq. (6) is a shape operator; an explicit embedding and derivation of the second fundamental form are missing. (3) Evaluation includes only random and entropy sampling with a single 5-NN classifier. Stronger instance-selection baselines, additional classifiers, and separate ablations of the score and adaptive threshold are needed to substantiate the claimed generality and methodological contribution.

Questions for authors

How can the method reduce annotation costs when its selection procedure already requires labels for the entire candidate pool? This supports labeled-data reduction, but not the claimed labeling-budget savings.

1 comment

Thank you for your careful reading of our paper and for the detailed feedback. We appreciate your positive remarks on the scoring procedure and on the experimental coverage across datasets, training fractions, and repetitions. Your comments are helpful, and we are working on the manuscript to implement any pertinent improvements.

Reviewer_Mz98OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce CuBAS, a method for selecting a training subset for supervised classification. They construct a kNN graph on the data and use Fisher information geometry to calculate the "curvature" of the resulting 1D statistical manifold. Using these curvature scores and an adaptive threshold selection algorithm, they split the data into low and high curvature points. They sample from both low and high curvature points at the same rate to form the training subset. They report that CuBAS outperforms random sampling and entropy-based sampling on 68 datasets from OpenML.

Critical strengths

In the reported experiments, the performance of CuBAS is consistently stronger than random sampling and entropy sampling. The method is simple and only adds O(nc) time after the kNN graph is built. The evaluation benchmark is quite broad and includes data from many different domains.

Critical weaknesses
  1. I'm not sure if the information-geometry curvature is measuring anything that we couldn't get from a simple baseline where a point's score is based on the heterogeneity of class labels in its neighborhood. For example, for a given query point, you could use neighborhood label entropy.

  2. I think the differential geometry is off in some places. It's not clear that this is actually "curvature" in a meaningful sense. Major: The authors state that the statistical manifold is 1-dimensional. If the manifold is 1-dimensional, then its intrinsic curvature is zero everywhere. What is the embedding space, if you're calculating extrinsic curvature? Minor: in the background, the given definition of the second fundamental form only makes sense for codimension 1.

Questions for authors

My understanding of algorithm 1 is that to obtain a partition into training and test data, you sample fraction p of the low-curvature nodes and fraction p of the high-curvature nodes for training. If that's the case, then you're using the same rate p for both low and high-curvature points, so the training and test sets have the same ratios of low/high curvature points. I would have thought that you'd want to sample low/high curvature points at different rates. If my understanding of algorithm 1 is correct, what is the mechanism for performance improvement? Can you clarify what exactly the actual code does?

1 comment

Thank you for taking the time to review our submission and for your thoughtful comments and questions. We appreciate your recognition of the method’s simplicity, the breadth of the experimental benchmark, and the consistency of the reported results. Your feedback is valuable to us. We are currently working on the paper and will incorporate any pertinent improvements in the revised version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.