acceptodds
Under review as a conference paper at ICLR 2027

Beyond Scalar Distances: Geometry-Aware Quantum Evidential C-Means via Contrastive Learning

Abstract

Traditional distance-based clustering methods lack the representational capacity for intricate uncertainty. To address this, evidential c-means (ECM) clustering introduces the concept of credal partition, providing a more flexible and expressive framework for uncertain data modeling. However, ECM fundamentally relies on a nondirectional scalar distance metric, focusing solely on how far samples are from centroids while ignoring their spatial distribution relative to overlapping clusters. To overcome these limitations, we we build upon generalized quantum evidence theory to propose a new method, quantum evidential c-means based on contrastive learning (CL-QECM). It not only integrates a fine-grained contrastive attraction-repulsion mechanism, but also encodes these spatial relationships into a geometry-aware angular representation within the complex credal space. Utilizing a temperature-scaled softmax function, the proposed method explicitly maximizes the similarity of positive pairs and minimizes the similarity of negative pairs in terms of distance. Extensive experiments on twelve benchmark datasets show that CL-QECM outperforms state-of-the-art methods in Purity, NMI, and ARI by up to 7.42%, 3.43%, and 7.30% respectively, exhibiting superior robustness and stability under noise and multi-cluster overlap.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Quantum evidential clustering, Contrastive learning, Quantum belief function theory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yUNHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CL‑QECM, an evidential clustering method that combines complex‑valued evidential assignment, distance‑induced phase mapping, and instance‑subset‑barycenter contrastive loss to tackle uncertainty and cluster overlap in clustering tasks.

Critical strengths
  1. The paper addresses the representation of ambiguous samples and overlapping clusters, which is a practically meaningful research problem.

  2. Experiments cover multiple types of datasets, report a variety of clustering metrics, and present ablation studies for both the contrastive‑loss module and the phase module.

Critical weaknesses
  1. Concerns remain regarding the incremental contribution of the proposed method. Evidential clustering assignments and prototype-based contrastive objectives have been explored in prior work, such as NN-EVCLUS: Neural Network-Based Evidential Clustering and Prototypical Cross-Domain Self-Supervised Learning for Few-Shot Unsupervised Domain Adaptation. While phase mapping could constitute the unique contribution of this paper, its current definition is still a deterministic function of distance. The paper fails to illustrate what geometric information it provides that cannot already be captured by distance alone.

  2. There appears to be an inconsistency between the phase formula and its core geometric interpretation. According to Equation (12), a shared normalization denominator is applied across distances from one sample to different subset barycenters. Consequently, equal distances from a sample to two subset barycenters will inevitably yield identical phases. Figure 2 illustrates the authors’ design intuition using distinct orientation deviations and boundary scales; however, it is not explicitly shown how these quantities enter Equation (12). Hence the claim that distinct phases arise under equal‑distance conditions in Figure 2 cannot be justified by the given formulation.

  3. There are inconsistencies between the objective function and update derivations as written. Equation (2) takes powers of complex‑valued evidential assignments, which generally produce complex numbers, yet Equations (8)‑(10) derive magnitude updates assuming a real‑valued objective built from powers of moduli. Furthermore, Equation (7) constructs the Lagrangian with the full joint objective, whereas subsequent derivatives drop the contrastive‑loss term. Although the main text mentions a surrogate objective for staged optimization, descent guarantees for the joint objective under this surrogate are not provided, and convergence proofs are missing.

  4. Baseline coverage and comparison settings need improvement. Several classical clustering methods are included, yet recent approaches are under‑represented. The authors should clarify their baseline‑selection criteria and adopt stronger baselines to enable valid assessment of performance gaps.

Questions for authors
  1. Could the authors provide convergence analysis for the proposed method and clarify whether the staged‑update scheme guarantees descent of the joint objective?

  2. Could the authors directly demonstrate the discriminative effect of phase on equidistant and overlapping‑cluster samples, alongside controlled comparisons against simple transformations that use identical distance information?

  3. UFCER has been introduced in Appendix G. Why is it absent from the comparison tables? Consequently, the most recent baseline included is from 2023, which fails to meet the requirement for adequate state‑of‑the‑art comparisons. Could the authors supplement additional stronger recent clustering baselines for evaluation?

Reviewer_QsMSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses unsupervised clustering with overlapping clusters and ambiguous boundaries. Conventional Evidential C Means (ECM) assigns data using only scalar Euclidean distances to cluster centers. When boundary points are equidistant to multiple centers, scalar distance fails to reflect their spatial layout. To solve this, the authors introduce two mechanisms into evidential clustering. First, real valued beliefs are extended to complex numbers with magnitude and phase to capture confidence and angular layout. Second, a contrastive loss between samples and subset barycenters is added to increase cluster separation during center updates. An alternating optimization scheme updates beliefs and centers iteratively. Experiments on twelve benchmark datasets evaluate clustering purity, NMI, and ARI.

Critical strengths
  1. When boundary samples are equidistant to multiple overlapping clusters, pure scalar distance cannot capture their spatial orientation. The proposed complex phase mapping provides a practical attempt to resolve this boundary ambiguity.
  2. Decoupling the nonlinear contrastive loss during belief updates allows a closed form solution via Lagrange multipliers, avoiding optimization instability from joint nonconvex updates.
  3. The method retains the empty set mechanism from classical evidence theory to absorb outliers and maps it into complex space, preserving robustness against isolated noise.
Critical weaknesses
  1. Imposing interclass repulsion during cluster center updates to prevent centers from sticking together is not a completely new concept. In fuzzy and possibilistic clustering, cluster repulsion methods like PCMR have long shared a similar goal of pulling competing cluster centers apart, and the mathematical form of expressing membership via polar magnitude and phase also directly parallels Complex Fuzzy C Means (CFCM). The manuscript would be strengthened by situating the proposed approach within this broader clustering literature and discussing how its repulsion mechanism and angular representation relate to and build upon these representative prior works.
  2. The construction of positive and negative pairs in this paper does not rely on data augmentations or multiview invariance priors commonly used in contrastive learning, but instead directly assigns positive and negative centers based on the model own belief assignments from the previous iteration. This mechanism is functionally closer to pseudolabel soft assignments in traditional EM clustering or self training. The conceptual basis for framing it as contrastive learning warrants further clarification, and the authors are advised to delineate the methodological attributes of this mechanism more rigorously.
  3. Appendix Table 5 shows that the algorithm experiences a noticeable surge in runtime on datasets with higher dimensions or more classes, such as the large runtime gap on USPS compared to traditional ECM. Considering that partitioning clustering typically offers high computational efficiency, repeatedly computing sample level contrastive gradients adds a notable operational burden. The authors are encouraged to provide a fuller discussion of limitations in the main text, clarifying the suitable scenarios and tradeoffs between performance gains and computational overhead.
Questions for authors
  1. The contrastive loss in Equation (5) essentially enforces interclass repulsion by computing Softmax negative distances across all samples. From a spatial geometry perspective, what fundamental differences and unique advantages does this sample to barycenter formulation offer in characterizing manifold distributions or theoretical properties, compared to classical approaches that directly constrain the relative positions of cluster centers?
  2. The phase angle in Equation (12) is defined as a deterministic algebraic mapping of normalized Euclidean distance. Given that the phase term is derived via an algebraic transformation of the scalar Euclidean distances already used to compute magnitudes, without introducing additional coordinate vectors or normal projections, how does this mapping break the constraint that pure scalar distances lack directionality mathematically or geometrically?
  3. The core claim of the paper is to break the nondirectional limitation of scalar distances using the phase angle. However, the ablation study shows that on several benchmark datasets, removing the phase does not lead to a performance drop and even yields slightly better metrics. Could the authors explain from a geometric distribution perspective why the angle term derived from Euclidean distance provides effective gains on some datasets, but appears insensitive or even redundant on others?
  4. The ablation study in Table 6 indicates that on multiple datasets such as PW, Australia, and USPS, the variant retaining only the early analytical update (w/o CL) already achieves clustering performance on par with or even slightly better than the full model. Given that the later momentum gradient refinement phase introduces a hundredfold increase in computational overhead, how do the authors evaluate the necessity and return on investment of introducing this later gradient refinement on datasets with low to moderate overlap or regular manifolds?
Reviewer_4EP8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CL-QECM, which combines evidential clustering, complex-valued uncertainty representations, and instance–centroid contrastive learning. The main motivation is that standard evidential c-means relies on distances to cluster centers, which may not adequately capture a point’s relationship with overlapping clusters. In evidential clustering, each point can assign belief to a subset of clusters, allowing uncertainty between multiple possible assignments to be represented explicitly. Belief assigned to the full set represents complete ignorance, while belief assigned to the empty set represents outliers. CL-QECM extends this representation using complex amplitudes and incorporates belief values into instance–centroid contrastive learning.

Critical strengths

The paper studies interesting problems and gives well motivation to the problem. Here are some strengths:

  1. The uncertainty-weighted contrastive formulation is reasonable. Using belief values to weight instance-centroid relationships makes sense, particularly for ambiguous points near cluster boundaries. I find this a useful way to incorporate uncertain assignments into contrastive learning.
  2. The empirical results are promising. The method achieves the best average ranks for Purity, NMI, and ARI among the compared methods. The paper also includes ablation studies to examine the contributions of individual components.
Critical weaknesses

While enjoying reading the paper, I have some concerns:

  1. The objective function needs clarification. Eq. (2) uses complex-valued (QβQ^\beta), whereas the magnitude updates appear to assume (∣Q∣β|Q|^\beta). It is therefore unclear which real-valued objective is being minimized and how phase affects its optimization. This may be a notation issue, but it is important for understanding the formulation.
  2. The claimed geometric contribution of phase is not sufficiently justified. Phase is computed from normalized distances. Converting these distances into angles does not, by itself, introduce directional information. I would like to see a clearer definition of the cluster-specific normalization and an explanation of how it supports the geometry claim.
  3. The ablation results are not fully convincing. On COIL20, removing phase alone causes a substantial drop, but removing both phase and contrastive learning largely recovers the performance. On USPS, removing both components gives slightly higher Purity than the full model. These results make the individual contributions difficult to assess and need further explanation.
Questions for authors

Please find the weaknesses part, with the following questions:

  1. What is the exact real-valued objective being minimized, and how does phase influence its optimization?
  2. Can a real-valued baseline using the same normalized-distance information achieve comparable performance?
  3. Were hyper-parameter tuning budgets and optimization settings matched across methods and ablation variants?
  4. Can the authors directly evaluate uncertainty quality and robustness under controlled levels of noise and cluster overlap?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.