acceptodds
Under review as a conference paper at ICLR 2027

DRP: Long-Tailed Deep Clustering via Denoising-Supported Repulsive Dirichlet Process

Abstract

In long-tailed deep clustering, dense head regions tend to occupy redundant prototype capacity, leading to head-class over-segmentation and tail-class absorption. To address this issue, we propose DRP, a nonparametric prototype inference framework for long-tailed deep clustering that rectifies prototype-capacity allocation bias by governing prototype birth, activation, repulsion and reallocation. DRP builds a semantic support measure via denoising consistency to stop dense regions from gaining unfair prototype birth advantages. Using this measure as the base distribution, it constructs a Dirichlet-process reference law and derives a Gibbs-tilted prototype prior minimizing KL divergence under a log-determinantal redundancy constraint. This prior controls active prototype count and suppresses redundant similar prototypes in head regions. For under-covered but stable semantic regions, cardinality-preserving prototype exchange reallocates redundant prototype slots. An identity-initialized residual adapter injects the corrected prototype structure back into the representation space; prototype-balanced feedback and conservative regularization limit representation drift. All components are integrated into a finite alternating MAP inference pipeline requiring no ground-truth labels, class frequencies or predefined class number. Extensive experiments validate DRP’s state-of-the-art performance on long-tailed clustering benchmarks, strong generalization to balanced data, and ability to infer unknown class counts.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Long-tailed Clustering, Bayesian Nonparametrics, Unsupervised learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_drKhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses long-tailed deep clustering, i.e., partitioning unlabeled data with a long-tailed class distribution without access to labels, class frequencies, or the number of classes. The authors argue that dense head regions occupy redundant prototype capacity, leading to head-class over-segmentation and tail-class absorption. The proposed DRP framework builds on a frozen encoder with a zero-initialized residual adapter and combines the following components, several of which are adapted from existing techniques:

  1. a denoising-based support score that acts as a one-sided barrier on where new prototypes can be born;
  2. a Gibbs-tilted Dirichlet-process prior with a log-determinant repulsion term to discourage redundant prototypes;
  3. a finite MAP objective with a prototype activation criterion;
  4. under-coverage candidate generation with cardinality-preserving prototype swapping;
  5. prototype-balanced adapter training.

The method is evaluated on CIFAR-10-LT, CIFAR-20-LT, and STL-10-LT (imbalance ratios 5 and 10), Tiny ImageNet-LT, and ImageNet-LT, as well as on balanced ImageNet subsets with a CLIP encoder. The paper also reports cluster-number estimation, component ablations, and hyperparameter sensitivity.

Critical strengths

1. Acceptable novelty through a coherent combination of components. The paper organizes prototype birth, activation, repulsion, and reallocation under a single structural objective for long-tailed clustering, rather than only re-weighting samples. Although several components are transferred from existing work, their integration for this problem is reasonable and constitutes an acceptable level of novelty.

2. Considerable accuracy improvements. DRP outperforms the strongest baseline (MiniClustering) on most long-tailed benchmarks. For example, on CIFAR-10-LT its ACC is higher by 4.5 points at IR = 5 and by 2.8 points at IR = 10 (Tab. 1–2). Gains are also reported on Tiny ImageNet-LT and ImageNet-LT. This is the main empirical support for the method.

Critical weaknesses

1. Insufficient rigor in experimental reporting. I consider the writing and reporting issues a bigger problem than the method itself, because they directly affect how much the comparisons can be trusted.

  • The baseline results in Tab. 1 and Tab. 2 appear to be taken from MiniClustering (ICLR 2026, https://proceedings.iclr.cc/paper_files/paper/2026/file/813ed01a11220fcdeed10e0d7da2de25-Paper-Conference.pdf). Most numbers are identical to Tab. 1 and Tab. 2 of that paper, but some individual entries do not match. For example, the STL-10 NMI of SeCu in Tab. 1 is 48.2 in MiniClustering but 46.0 in this paper.
  • Line 408 states that a frozen CLIP ViT-B/16 image encoder is used "for all compared variants so that representation quality is controlled". However, the TAC results in Tab. 5 appear to be copied directly from the original TAC paper (https://arxiv.org/html/2310.11989v3), which uses a CLIP ViT-B/32 encoder. The comparison in Tab. 5 is therefore not controlled as claimed.

2. Heavy formula presentation and a bloated design.

  • The method contains many components, which leads to a large number of redundant hyperparameters and makes the framework somewhat bloated.
  • The paper also stacks many theoretical formulas, which hurts readability. Derivations that are already consensus in the field (e.g., standard variational or Gibbs-type results) could be abbreviated. For the necessary ones, I suggest naming common structures before referring to them and using underbrace annotations. For example, Eq. 11 appears to contain the negative free energy of EM for a vMF mixture model; annotating it as such would help readers see which theoretical sources the paper builds on.
  • There is also a notational error at Line 296: "C_r = c̃_m ∈ I_r" equates a set (C_r) with an element (c̃_m), which cannot hold. I believe this is a typo.
Questions for authors

Clarifications:

  1. Are the baseline numbers in Tab. 1 and Tab. 2 reproduced by the authors or taken from MiniClustering? If they are taken from MiniClustering, are the mismatched entries (e.g., SeCu on STL-10, NMI: 48.2 vs. 46.0) typos introduced during transcription? Please clarify the source of each baseline result.
  2. Regarding the statement at Line 408: were the TAC results in Tab. 5 obtained with CLIP ViT-B/16, or taken from the original paper (ViT-B/32)? To keep the paper rigorous, please either re-run the baselines with ViT-B/16 or remove this statement.
  3. At Line 296, should the definition of C_r be {c̃_m : m ∈ I_r} (consistent with Appendix Eq. 40)? Please correct it.

Suggestions:

  1. Consider abbreviating derivations that are standard in the field, and annotating necessary formulas with underbraces and named structures (e.g., the vMF mixture EM free energy in Eq. 11) to improve readability.

If the authors resolve these issues, I will consider raising my score.

5 comments

Thank you for your careful reading of our manuscript, detailed feedback, and constructive suggestions. Below, we respond to each of your comments and explain the corrections we have made and the additional supporting evidence we provide.

[W1] Baseline Reporting and Backbone Consistency.

[W1-1] Sources and Corrections of Baseline Results.

We agree with the reviewer that rigorous experimental reporting is essential for assessing the reliability of the comparisons. We followed the MiniClustering’s experimental settings and reproduced our experiments under the same protocol. For most competing methods, we report the results provided by the MiniClustering to ensure a consistent comparison. We acknowledge the two incorrect entries in our tables and sincerely apologize for these mistakes. We have corrected them in the revised version and carefully rechecked all reported results. To facilitate verification and reproducibility, we will release our code and experimental configurations.

We have corrected the results in the revised version paper (Table 1) and carefully rechecked all reported results.

[W1-2] Backbone Settings for the Comparison with TAC.

We thank the reviewer for this careful observation. The TAC results reported in the original submission use a CLIP ViT-B/32 encoder, whereas our initial DRP experiments used a frozen CLIP ViT-B/16 encoder. We have revised the statement at Line 402, additionally evaluated DRP with ViT-B/32, and included these results in Table 5 of the revised manuscript. As shown in the Table below, DRP outperforms TAC across all reported metrics on all four ImageNet datasets when using the same backbone architecture, demonstrating DRP’s effectiveness with both backbone configurations.

Table 12. Balanced ImageNet clustering (%).

DatasetMethodACCNMIARI
ImageNet-50TAC (ViT-B/32)78.783.767.5
ImageNet-50DRP (ViT-B/32)83.192.278.3
ImageNet-50DRP (ViT-B/16)84.391.879.7
ImageNet-100TAC (ViT-B/32)63.868.157.6
ImageNet-100DRP (ViT-B/32)77.589.372.4
ImageNet-100DRP (ViT-B/16)78.189.071.2
ImageNet-200TAC (ViT-B/32)59.270.950.8
ImageNet-200DRP (ViT-B/32)74.486.565.2
ImageNet-200DRP (ViT-B/16)74.887.666.3
ImageNet-1KTAC (ViT-B/32)58.279.943.5
ImageNet-1KDRP (ViT-B/32)60.182.343.9
ImageNet-1KDRP (ViT-B/16)59.381.744.7

We have added DRP results obtained with ViT-B/32 to Table 5 of the revised manuscript and refined the description of DRP.

[W2] Framework Complexity, Formula Readability, and Notation.

[W2-1] Component Contributions, Computational Cost, and Parameter Sensitivity.

Thank you for raising this concern. The components of DRP address different aspects of prototype allocation. Denoising support filters unsupported candidate prototypes, Gibbs repulsion suppresses redundant prototypes, under-coverage candidate generation identifies regions requiring additional capacity, and prototype-balanced adaptation balances the contributions of individual prototypes to representation learning. The ablation results in Table 7 of the paper support their contributions, with the removal of individual components reducing ACC on CIFAR-10-LT by 1.3–2.5 percentage points. Appendix C.4 further analyzes the computational cost. The denoiser is trained only once, and the repulsion term operates on a K+×K+K_+\times K_+ matrix, with its cubic cost depending on the number of active prototypes K+K_+. With a fixed feature dimension, prototype budget, and number of optimization rounds, the overall computational cost scales approximately linearly with the number of samples. We further compared the runtimes of DRP and MiniClustering on CIFAR-10-LT (IR = 10), excluding BYOL pretraining for both methods. DRP takes approximately 28.33 minutes, compared with 26.8 minutes for MiniClustering, corresponding to an additional 1.53 minutes, or approximately 5.7%. Together with the ACC improvement from 64.6% to 67.4% reported in Table 2 of the paper, this comparison indicates that DRP achieves better performance with a modest increase in runtime. Although DRP introduces additional hyperparameters compared with MiniClustering, Figure 3 and Table 10 of the paper show that its performance is relatively insensitive to moderate variations in τs\tau_s, τu\tau_u, and κmax⁡\kappa_{\max}. This suggests that the tuning effort remains manageable, allowing DRP to achieve substantial performance improvements with a modest increase in runtime.

[W2-2] Streamlined Derivations and Annotated Equations.

Thank you for these suggestions on improving readability. We have streamlined the standard derivations and clarified the structure of the main objective. We have explicitly labeled the fitting term in Eq. (11) as the EM free energy of a vMF mixture model, combined the fitting terms into a single expression, and used underbraces to distinguish the EM free energy, activation cost, support barrier, and prototype repulsion terms. We have also added explanatory text to improve readability. The revised Eq. (11) is shown below:

\mathcal F_{\mathrm{struct}} ={}& \underbrace{-\sum_i\sum_{m\in\mathcal I_r} q_{im}\left(\log\pi_m+\kappa z_i^\top\tilde c_m-\log q_{im}\right)}_{\mathcal E_{\mathrm{fit}}:\\ \text{vMF-mixture EM free energy}} \\\\ &+\underbrace{\lambda_a^{\mathrm{eff}}\lVert r\rVert_0}_{\text{activation cost}}+\underbrace{\lambda_s\sum_m r_m b_\psi(\tilde c_m)}_{\text{support barrier}} \\\\ &+\underbrace{\lambda_g\mathcal E_{\mathrm{rep}}(\mathcal C_r)}_{\text{prototype repulsion}}. \end{aligned}$$ We have abbreviated the standard variational steps in Sec. 3.2 and 3.3. Specifically, we moved the base-measure variational formulation in the original Eq. (3) to Appendix B.6 and replaced its main-text presentation with a concise description. We similarly refer readers to Appendix A.2 for the intermediate change-of-measure expression in the original Eq. (8) and its derivation. The main text now explicitly invokes the standard Gibbs variational principle and directly presents the resulting exponentially tilted base measure and Gibbs-tilted DP prior. We have also added underbrace annotations to the original Eq. (9) to identify their component structures. In the original Eq. (9), the annotations distinguish the CRP partition prior, denoising-support factor, and prototype-repulsion factor. We have revised the accompanying text to explain how these factors govern the partition, individual prototype locations, and their joint geometry, and how the repulsion factor contributes to the MAP objective. The annotated version of the original Eq. (9) is shown below: $$p_{\mathrm G}(\Pi_N,C\mid\psi)=\frac{1}{\mathcal Z_{G,N}}\underbrace{p_{\mathrm{CRP}}(\Pi_N\mid\alpha)}_{\text{CRP partition prior}}\\,\underbrace{\prod\nolimits_{k=1}^{K_+}h_\psi(c_k)}_{\text{denoising support}}\\,\underbrace{\det(\bar L_C)^{\lambda_g}}_{\text{prototype repulsion}}.$$ **We have refined the manuscript by streamlining the presentation of some standard derivations and adding explanatory annotations to the equations to improve readability.** ## [W2-3] Correction of the Active Prototype Set Notation. We thank the reviewer for pointing this out. We have corrected the definition of $C_r$ (now line 291 of the revised manuscript).

[Q1] Clarification of Baseline Result Sources.

As clarified in our response to the first point of [W1-1] above, the baseline results in Tables 1 and 2 of the paper come from two sources. We reproduced the DGD results, while all other baseline results were taken from the MiniClustering paper. The discrepancy in the SeCu entry was due to a transcription error, and this entry has been corrected in the revised manuscript.

[Q2] Backbone Alignment for the Comparison with TAC.

As clarified in our response to [W1-2] above, the TAC results were taken from the original TAC paper and use CLIP ViT-B/32, whereas our initial DRP results used ViT-B/16. We additionally evaluated DRP with ViT-B/32 to provide a comparison using the same backbone architecture. Under this setting, DRP outperforms TAC across all reported metrics on all four balanced ImageNet datasets, as shown in the results provided in our response to [W1-2].

[Q3] Correction of the Definition of CrC_r.

We thank the reviewer for pointing this out. We have corrected the definition of CrC_r on line 296.

[Q4] Improved Readability of Derivations and Equations.

Thank you for these constructive suggestions. As detailed in our response to [W2-2] above, we have abbreviated the standard variational and Gibbs derivations in the main text and referred readers to the appendices for intermediate steps. We have also explicitly identified the vMF-mixture EM free-energy term and added underbraces to distinguish the fitting, activation, support, and prototype-repulsion components. The annotated equations and accompanying explanations are provided in that response.

We have streamlined the presentation of standard derivations (e.g., lines 223–226 of the revised manuscript) and refined the presentation of the equations (e.g., Eqs. 7 and 9 in the revised manuscript). We have also revised passages in the main text for clarity (e.g., lines 292–295 in the revised manuscript) and improved the exposition in the appendix (e.g., Sections A.1 and B.6 in the revised manuscript).

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

The paper proposes DRP for long-tailed deep clustering with an unknown number of classes. It identifies prototype-capacity allocation bias, stating that dense head classes can occupy redundant prototypes while sparse tail classes remain insufficiently covered. DRP combines a denoising-supported base measure, log-determinantal prototype repulsion, and cardinality-preserving prototype exchange. A residual adapter incorporates the resulting clustering structure through prototype-balanced updates and conservative regularization. Experiments on several long-tailed and balanced image datasets show improved clustering performance, complemented by cluster-count estimation, component ablations, sensitivity analyses, and additional evaluations across random seeds and representation-learning frameworks.

Critical strengths
  1. The framework provides a coherent mechanism for correcting prototype-allocation bias. All the designed components address complementary aspects of the problem: denoising support discourages unsupported prototype locations, repulsion penalizes redundant configurations, and prototype exchange reallocates capacity to under-covered regions. The cardinality-preserving exchange is particularly useful because it allows a candidate to replace an inefficient prototype without paying the additional fixed cost of increasing the cluster count. Together, these mechanisms provide a concrete approach to improving tail coverage. And it performs competitively across multiple long-tailed datasets, indicating the robustness and the effectiveness of the approach.

  2. The paper provides useful theoretical grounding for its main components. The support measure and Gibbs-tilted prior are derived through KL-regularized variational formulations. The appendix also characterizes prototype activation and cardinality-preserving exchange, including conditions for structural-objective descent with fixed representations. These analyses make the roles of the proposed components explicit and strengthen the rationale behind the framework.

Critical weaknesses
  1. The paper does not sufficiently explain how the candidate inventory size and activation-related parameters are selected without using labels or the true class count. Table 9 shows that changing Kmax⁡K_{\max} from 250 to 400 changes the estimated cardinality from 164.2 to 215.6, making this selection procedure consequential. A comparison with an established unknown-cardinality method such as DeepDPM, using matched representations, would also help establish the specific advantage of DRP’s cardinality inference.

  2. The construction of under-coverage candidates is insufficiently specified. Equation 13 forms each candidate by aggregating samples within a local neighborhood Nm\mathcal{N}_m, but does not explain how these neighborhoods are constructed, how their anchor samples are selected, or how their sizes are determined. Large neighborhoods may mix nearby semantic classes, while small neighborhoods may produce noisy or redundant candidates. A more explicit procedure would improve reproducibility and clarify how the method identifies under-covered tail regions.

Questions for authors
  1. Please explain how Kmax⁡K_{\max} and the activation-related parameters are selected without labels or the true class count.

  2. Please explain how the neighborhoods Nm\mathcal{N}_m in Equation 13 are constructed, including anchor selection, neighborhood size, and any filtering of overlapping or redundant proposals. Pseudocode and the corresponding parameter settings would help.

  3. Future work: A comparison with an established unknown-cardinality method such as DeepDPM using matched representations, would help establish the specific advantage of DRP’s cardinality inference.

5 comments

Thank you for your thoughtful review and constructive suggestions. We carefully consider your concerns below and explain the corresponding clarifications and revisions to the manuscript.

[W1] Candidate Capacity, Parameter Selection, and Unknown-Cardinality Clustering.

We address parameter selection, sensitivity to the candidate inventory size, and comparison with an unknown-cardinality baseline below.

1. Selecting the inventory size and activation-related parameters without labels.

We use internal clustering metrics, primarily the cosine silhouette score, together with the empirical diagnostics described below, to identify candidate parameter ranges from trends and stable regions in the evaluation curves. These metrics require no labeled samples. As shown in Table 8, the near-optimal ranges overlap substantially across the reference settings, with stable internal quality and partition among the accepted configurations. We therefore reuse the shared ranges as the default search ranges for subsequent datasets, reducing the need to restart tuning over a broad domain. For a new dataset, we first examine values within the default ranges and perform additional local checks when its size, dimensionality, or noise characteristics differ substantially from those of the reference datasets. Activation and pruning records help identify overly conservative activation, prototype redundancy, or insufficient support, and guide the direction of refinement. For Kmax⁡K_{\max}, we conduct a separate capacity search for each dataset using internal evaluation metrics. Among candidates with comparable internal quality and stable predictions, we favor the smaller capacity.

Table 8. Near-optimal coefficient intervals and ranges of ARI among the selected configurations.

SettingNear-optimal λaeff\lambda_a^{\mathrm{eff}} intervalNear-optimal λs\lambda_s intervalNear-optimal λg\lambda_g intervalARI
CIFAR-10-LT, IR = 530–50400–100010–200.627–0.632
CIFAR-10-LT, IR = 1030–50400–100010–200.609–0.611
STL-10-LT, IR = 530–50600–100010–200.433–0.445
STL-10-LT, IR = 1030–55400–9008–200.395–0.403
Supported Ranges30–50600–90010–20—

2. Sensitivity To Kmax⁡K_{\max} and Comparison With DeepDPM.

The results in Table 9 of the original paper show that varying Kmax⁡K_{\max} can lead to errors in the estimated number of clusters. This reflects a challenge in unsupervised learning (Such as DeepDPM, DBSCAN and moVB), where cluster-number estimates can deviate substantially from the true value in some settings. Nevertheless, as shown in Table 9, DRP’s mechanism for correcting bias in prototype capacity allocation enables it to consistently achieve better clustering performance than methods that assume a known number of clusters, such as MiniClustering. Using the same representations, we also compared DRP with DeepDPM, a method that supports an unknown number of clusters. As shown in Table 10 below, DRP outperforms DeepDPM in both clustering performance and the accuracy of cluster-number estimation.

Table 9. Sensitivity to the truncation level Kmax⁡K_{\max}.

Kmax⁡K_{\max}ACCCAANMIARI
25027.021.542.816.5
30026.721.743.117.3
32027.122.343.417.3
40025.521.842.917.5
MiniClustering25.519.440.614.9

Table 10. Clustering performance of DeepDPM and our method (DRP) on long-tailed datasets.

DatasetMethodACCCAANMIARIInferred Clusters
CIFAR-10-LTDeepDPM45.252.856.739.218.0
CIFAR-10-LTDRP78.876.269.963.213.0
CIFAR-20-LTDeepDPM35.836.238.318.631.0
CIFAR-20-LTDRP48.344.449.333.222.0
STL-10-LTDeepDPM50.145.745.238.316.0
STL-10-LTDRP57.654.555.044.110.0

We have added The Parameter Selection Method in Appendix C.5, The DeepDPM Comparison Experiment in Appendix D.1.

[W2] Construction of Under-Coverage Candidates.

We construct Nm\mathcal N_m as a local neighborhood centered on an under-covered sample, subject to a maximum cosine distance from the anchor.

1. Anchor selection We compute the sample weights and define the eligible set as

ωiuc=[1−max⁡k∈Irzi⊤c~k−τu]+1[sψ(ui)≥τs],E=i:ωiuc>0.\omega_i^{uc} = \left[1-\max_{k\in I_r}z_i^\top\tilde c_k-\tau_u\right]_+ \mathbf{1}[s_\psi(u_i)\ge\tau_s], \qquad \mathcal E=\\{i:\omega_i^{uc} > 0\\}.

We select the sample with the largest ωiuc\omega_i^{uc} as the first anchor. We then select each subsequent anchor greedily according to

am=arg⁡max⁡i∈E∖Aωiuc(10−3+min⁡a∈A(1−ui⊤ua)).a_m = \arg\max_{i\in\mathcal E\setminus\mathcal A} \omega_i^{uc} \left(10^{-3}+\min_{a\in\mathcal A}(1-u_i^\top u_a)\right).

This rule favors samples with higher under-coverage weights and larger distances to their nearest selected anchor, reducing the tendency for anchors to concentrate in the same local region.

2. Neighborhood construction In the fixed feature space UU, we query the kmax⁡−1k_{\max}-1 nearest neighbors of anchor ama_m by cosine distance over the full dataset, excluding the anchor itself. We then add the anchor and retain only samples with positive weights whose cosine distance to the anchor does not exceed rlocr_{\mathrm{loc}}. The resulting neighborhood is

\mathcal N_m = \left(\\{a_m\\}\cup\operatorname{kNN}^{U}_{k_{\max}-1}(a_m)\right) \cap\mathcal E \cap\left\\{i:1-u_i^\top u_{a_m}\le r_{\mathrm{loc}}\right\\}.

The cap kmax⁡k_{\max} controls the number of samples used for aggregation, while rlocr_{\mathrm{loc}} limits the neighborhood’s geometric extent and reduces the risk of mixing samples from nearby semantic regions.

3. Proposal deduplication We aggregate the current adapted features ZϕZ_\phi according to Eq. (13). For each new proposal mm and previously retained proposal nn, we compute the cosine similarity SS between their directions and the Jaccard overlap of their neighborhoods, J=∣Nm∩Nn∣/∣Nm∪Nn∣J=|\mathcal N_m\cap\mathcal N_n|/|\mathcal N_m\cup\mathcal N_n|. If the following condition holds, we treat the two proposals as duplicates and retain the one processed first:

Dup⁡(S,J)=[S≥0.995]∨[S≥0.98∧J≥0.80].\operatorname{Dup}(S,J) = [S\ge 0.995]\\ \lor\\ [S\ge 0.98\\ \land\\ J\ge 0.80].

4. Pseudocode and Parameter Settings. We further present the pseudocode in Appendix A.5 and provide the corresponding parameter settings as follows. We set the relative normalization tolerance to ε=10−7\varepsilon=10^{-7}, the maximum neighborhood size to kmax⁡=35k_{\max}=35, and the maximum cosine distance from the anchor to rloc=0.25r_{\mathrm{loc}}=0.25. We treat two proposals as duplicates if S≥0.995S\ge0.995 or if both S≥0.98S\ge0.98 and J≥0.80J\ge0.80.

We have added the under-coverage candidate construction method in Appendix A.5, including the pseudocode (Algorithm 1).

[Q1] Label-Free Selection of Candidate Capacity and Activation Parameters.

Our procedure is as follows. We use internal clustering metrics, primarily the cosine silhouette score, together with the empirical diagnostics described below, to identify candidate parameter ranges from trends and stable regions in the evaluation curves. These metrics require no labeled samples. As shown in Table 11, the near-optimal ranges overlap substantially across the reference settings, with stable internal quality and partition among the accepted configurations. We therefore reuse the shared ranges as the default search ranges for subsequent datasets, reducing the need to restart tuning over a broad domain. For a new dataset, we first examine values within the default ranges and perform additional local checks when its size, dimensionality, or noise characteristics differ substantially from those of the reference datasets. Activation and pruning records help identify overly conservative activation, prototype redundancy, or insufficient support, and guide the direction of refinement. For Kmax⁡K_{\max}, we conduct a separate capacity search for each dataset using internal evaluation metrics. Among candidates with comparable internal quality and stable predictions, we favor the smaller capacity.

Table 11. Near-optimal coefficient intervals and ranges of ARI among the selected configurations.

SettingNear-optimal λaeff\lambda_a^{\mathrm{eff}} intervalNear-optimal λs\lambda_s intervalNear-optimal λg\lambda_g intervalARI
CIFAR-10-LT, IR = 530–50400–100010–200.627–0.632
CIFAR-10-LT, IR = 1030–50400–100010–200.609–0.611
STL-10-LT, IR = 530–50600–100010–200.433–0.445
STL-10-LT, IR = 1030–55400–9008–200.395–0.403
Supported Ranges30–50600–90010–20—

We have added The Parameter Selection Method in Appendix C.5.

[Q2] Anchor Selection, Neighborhood Construction, and Proposal Deduplication.

As detailed in our response to [W2] above, anchors are selected from supported, under-covered samples according to their under-coverage weights and distances from previously selected anchors. For each anchor, we query cosine nearest neighbors in the fixed feature space and retain eligible samples within a cosine distance of rloc=0.25r_{\mathrm{loc}}=0.25, with the neighborhood size capped at kmax⁡=35k_{\max}=35. Candidate directions are formed by under-coverage-weighted aggregation of the current adapted features, and redundant proposals are filtered using directional similarity and neighborhood overlap. The detailed selection rule, screening criteria, and pseudocode are provided in that response.

We have added the under-coverage candidate construction method in Appendix A.5, including the pseudocode (Algorithm 1).

[Q3] Comparison with DeepDPM Using Matched Representations.

Thank you for your constructive suggestion. As reported in our response to [W1] above, we compared DRP with DeepDPM using the same input representations on CIFAR-10-LT, CIFAR-20-LT, and STL-10-LT at IR$=5$. DRP achieves higher ACC, CAA, NMI, and ARI on all three datasets, with reported cluster-count estimates closer to the ground-truth values. The complete results are provided in Table 10 in our response to [W1].

We have added the DeepDPM Comparison Experiment in Appendix D.1.

Reviewer_8GmUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DRP, a long-tailed deep clustering framework that addresses prototype-capacity allocation bias. It combines denoising-based semantic support, Gibbs repulsion, prototype reallocation, and prototype-balanced representation adaptation to reduce head-class over-segmentation, improve tail-class coverage, and infer the number of clusters.

Critical strengths
  1. The paper identifies a critical problem in long-tailed clustering: redundant prototype allocation to head classes and insufficient coverage of tail classes.
  2. The proposed framework provides a coherent integration of prototype support, redundancy suppression, and dynamic reallocation.
  3. Extensive experiments on long-tailed and balanced benchmarks show consistent improvements over competitive baselines.
Critical weaknesses
  1. Key coefficients, particularly λaeff\lambda_a^{\mathrm{eff}}, λs\lambda_s, and λg\lambda_g, are not systematically evaluated. Their effects on clustering performance and estimated cluster cardinality across different imbalance ratios remain unclear.
  2. Although the method is motivated by head-class prototype redundancy and tail-class under-coverage, the experiments mainly report aggregate clustering metrics. More direct analyses of prototype allocation across head and tail classes, particularly under different imbalance ratios, are needed to establish whether the improvements arise from the proposed reallocation mechanism.
  3. The proposed Dirichlet-process prior is ultimately approximated by a finite MAP objective with a linear cardinality penalty. The practical benefit of this construction over simpler regularized prototype clustering objectives is not clearly established.
Questions for authors
  1. How are λaeff\lambda_a^{\mathrm{eff}}, λs\lambda_s, and λg\lambda_g selected without access to the true class count or class frequencies? Could the authors report sensitivity analyses of clustering performance and estimated KK under different imbalance ratios?
  2. Can the authors provide quantitative analyses of head/tail prototype allocation and tail-class coverage before and after DRP, across different imbalance ratios? Additionally, could they isolate the contribution of prototype swapping from under-coverage candidate generation?
  3. What practical advantage does the Gibbs-tilted Dirichlet-process formulation offer over a simpler finite prototype objective with cardinality and repulsion regularization?
4 comments

Thank you for your insightful assessment and constructive feedback. We address each weakness and your question below, with the relevant evidence and manuscript locations.

[W1] Sensitivity of Key Coefficients Across Imbalance Ratios.

We conducted additional experiments to evaluate the effects of λaeff\lambda_a^{\mathrm{eff}}, λs\lambda_s, and λg\lambda_g on CIFAR-10-LT, using imbalance ratios of 5 and 10, adopting a one-variable-at-a-time approach where only the target hyperparameter is varied while all others remain fixed and consistent with those used in the main results. The results are reported in Tables 1-3 below.

  1. For λaeff\lambda_a^{\mathrm{eff}}. The results in Table 1 show that increasing λaeff\lambda_a^{\mathrm{eff}} from 25 to 55 reduces the estimated number of clusters at both IR = 5 and IR = 10. Clustering performance remains relatively stable over the range of 30–50 under both imbalance settings, but declines when λaeff\lambda_a^{\mathrm{eff}} is further increased to 55. The decrease in CAA is more pronounced at IR = 10 than at IR = 5.

Table 1. Sensitivity to λaeff\lambda_a^{\mathrm{eff}} on CIFAR-10-LT, with λs=800\lambda_s=800 and λg=15\lambda_g=15 fixed.

IRλaeff\lambda_a^{\mathrm{eff}}ACCCAANMIARIK^\widehat{K}
52578.075.570.062.414.2
53078.576.070.263.013.5
54078.876.269.963.213.0
55078.675.869.863.112.7
55578.275.269.562.612.4
102566.763.465.760.412.7
103067.263.866.061.012.1
104067.463.965.761.111.7
105067.263.565.560.911.4
105566.662.765.160.311.2
  1. For λs\lambda_s. The results in Table 2 show that both clustering performance and cluster-number estimates remain relatively stable, with only minor fluctuations, as λs\lambda_s increases from 400 to 1000.

Table 2. Sensitivity to λs\lambda_s on CIFAR-10-LT, with λaeff=40\lambda_a^{\mathrm{eff}}=40 and λg=15\lambda_g=15 fixed.

IRλs\lambda_sACCCAANMIARIK^\widehat{K}
540078.475.969.862.813.3
560078.776.169.863.013.3
580078.876.269.963.213.0
590078.676.269.963.113.2
5100078.476.069.662.713.5
1040067.063.465.761.011.9
1060067.163.765.861.111.9
1080067.463.965.761.111.7
1090067.363.865.760.911.8
10100067.363.665.660.912.0
  1. For λg\lambda_g. The results in Table 3 show that increasing λg\lambda_g from 8 to 22 reduces the estimated number of clusters at both IR = 5 and IR = 10. Under both imbalance settings, the mean values of all four clustering metrics vary by no more than 0.8 percentage points. Within the range of 10–20, each metric varies by no more than 0.3 percentage points, indicating stable clustering quality near the reference configuration, whereas further increasing the weight leads to slight declines in some metrics.

Table 3. Sensitivity to λg\lambda_g on CIFAR-10-LT, with λaeff=40\lambda_a^{\mathrm{eff}}=40 and λs=800\lambda_s=800 fixed.

IRλg\lambda_gACCCAANMIARIK^\widehat{K}
5878.275.769.562.713.6
51078.676.069.863.013.3
51578.876.269.963.213.0
52078.776.369.963.112.8
52278.475.969.762.812.7
10866.863.265.360.612.3
101067.263.765.661.012.0
101567.463.965.761.111.7
102067.364.065.861.011.6
102267.063.565.460.711.5

We have added sensitivity analyses of λaeff\lambda_a^{\mathrm{eff}}, λs\lambda_s, and λg\lambda_g in Sec. 4.3.

[W2] Head/Tail Prototype Allocation and Tail-Class Coverage.

We conducted an additional quantitative analysis on CIFAR-10-LT and CIFAR-20-LT at imbalance ratios of 5 and 10. We compared prototype allocation for head and tail classes and tail-class coverage before and after applying DRP. The results are shown in Table 4.

Table 4. Prototype allocation and tail-class coverage before and after applying DRP.

DatasetIRHead-dominant prototypes / BeforeHead-dominant prototypes / AfterTail-dominant prototypes / BeforeTail-dominant prototypes / AfterTail-class coverage (%) / BeforeTail-class coverage (%) / After
CIFAR-10-LT5541233.366.7
CIFAR-10-LT10651233.366.7
CIFAR-20-LT511102333.350.0
CIFAR-20-LT101091216.733.3

The results show that DRP reduces the number of redundant prototypes assigned to head classes, allocates additional prototypes to tail classes, and improves tail-class coverage.

We have added the prototype allocation experiment in Appendix D.2.

[W3] Finite MAP Approximation and Prototype Regularization.

DRP uses a finite MAP surrogate for practical inference. The DP partition term is approximated by a linear cardinality penalty, while the denoising-support barrier and the Gibbs-derived log-determinant repulsion are retained. The probabilistic construction jointly describes adaptive cardinality, prototype support, and configuration-level redundancy. With the representation fixed, prototype birth, pruning, and swapping are evaluated through the same structural objective, balancing fitting gains against activation, support, and redundancy costs.

The experiments assess the practical contribution of these retained terms. In Table 7 of the manuscript, adding Gibbs repulsion on CIFAR-10-LT with IR$=10$ improves ACC from 65.5% to 67.4% and NMI from 63.8% to 65.7%. We further compare log-det repulsion with pairwise repulsion while retaining the same data-fitting term and cardinality penalty. As shown in the comparison below, log-det repulsion improves ACC and NMI by 1.2 and 1.9 percentage points, respectively, and increases tail-sample coverage from 26.9% to 30.9%. These results support the practical value of the set-level repulsion used in the finite surrogate. The DP reference and Gibbs tilting provide a unified probabilistic interpretation of this objective, rather than a separate exact DP inference procedure.

Table 5. Comparison of pairwise and log-det repulsion on CIFAR-10-LT (IR = 10).

SettingACCNMICAAARIK^\hat{K}Tail-sample coverage (%)
pairwise66.263.861.959.514.226.9
log-det67.465.763.961.111.730.9

We have added the experiment comparing repulsion formulations in Appendix D.3.

[Q1] Label-Free Parameter Selection and Sensitivity Analysis.

I. Parameter selection in an unsupervised setting

Our procedure is as follows. We use internal clustering metrics, primarily the cosine silhouette score, together with the empirical diagnostics described below, to identify candidate parameter ranges from trends and stable regions in the evaluation curves. These metrics require no labeled samples. As shown in Table 6, the near-optimal ranges overlap substantially across the reference settings, with stable internal quality and partition among the accepted configurations. We therefore reuse the shared ranges as the default search ranges for subsequent datasets, reducing the need to restart tuning over a broad domain. For a new dataset, we first examine values within the default ranges and perform additional local checks when its size, dimensionality, or noise characteristics differ substantially from those of the reference datasets. Activation and pruning records help identify overly conservative activation, prototype redundancy, or insufficient support, and guide the direction of refinement.

Table 6. Near-optimal coefficient intervals and ranges of ARI among the selected configurations.

SettingNear-optimal λaeff\lambda_a^{\mathrm{eff}} intervalNear-optimal λs\lambda_s intervalNear-optimal λg\lambda_g intervalARI
CIFAR-10-LT, IR = 530–50400–100010–200.627–0.632
CIFAR-10-LT, IR = 1030–50400–100010–200.609–0.611
STL-10-LT, IR = 530–50600–100010–200.433–0.445
STL-10-LT, IR = 1030–55400–9008–200.395–0.403
Supported Ranges30–50600–90010–20—

II. Additional hyperparameter experiments

As detailed in our response to [W1] above, we conducted one-variable-at-a-time sensitivity experiments for λaeff\lambda_a^{\mathrm{eff}}, λs\lambda_s, and λg\lambda_g on CIFAR-10-LT at imbalance ratios of 5 and 10, evaluating both clustering performance and the estimated number of clusters. Tables 1–3 show that increasing λaeff\lambda_a^{\mathrm{eff}} or λg\lambda_g reduces the estimated cluster count, whereas varying λs\lambda_s produces relatively small fluctuations in both performance and cardinality. Clustering performance remains relatively stable over moderate parameter ranges under both imbalance settings, but declines when λaeff\lambda_a^{\mathrm{eff}} is increased to 55.

We have added the Parameter Selection Method in Appendix C.5, Hyperparameter Experiments in Sec. 4.3.

[Q2] Prototype Allocation and Ablation of Candidate Generation and Swapping.

1. Head/tail prototype allocation and tail-class coverage. As discussed in our response to [W2] above, we quantitatively analyzed prototype allocation and tail-class coverage on CIFAR-10-LT and CIFAR-20-LT at imbalance ratios of 5 and 10. Across these settings, DRP reduces the number of head-dominant prototypes, increases the number of tail-dominant prototypes, and improves tail-class coverage. The detailed results and analysis are provided in that response.

2. Ablation of under-coverage candidate generation and prototype swapping. We further evaluate their contributions on CIFAR-10-LT (IR$=10$) by first enabling candidate generation and then adding prototype swapping. As shown in Table 7, enabling candidate generation without swapping improves ACC, NMI, CAA, and ARI by 1.3, 0.7, 0.9, and 1.5 percentage points, respectively, compared with disabling both components. With candidate generation enabled, adding prototype swapping further improves ACC, NMI, CAA, and ARI by 1.2, 1.1, 1.4, and 1.2 percentage points, respectively, isolating the additional benefit of swapping beyond candidate generation alone. The improvements across all four metrics support the benefit of combining candidate discovery with prototype-capacity reallocation.

Table 7. Ablation of under-coverage candidate generation and prototype swapping on CIFAR-10-LT with IR=10\mathrm{IR}=10.

SettingUnder-coverage candidatesPrototype swappingACCNMICAAARI
Both Off64.963.961.658.4
Candidates Only✓\checkmark66.264.662.559.9
Full✓\checkmark✓\checkmark67.465.763.961.1

We have added the Ablation Studies in Appendix D.4.

[Q3] Practical Benefits of the Gibbs-Tilted DP Formulation.

Thank you for your question. As detailed in our response to [W3] above, the formulation provides a unified probabilistic construction of adaptive cardinality, denoising support, and set-level prototype repulsion. Its practical implementation is the finite MAP surrogate, which evaluates prototype birth, pruning, and swapping through a common structural objective. The comparison reported in that response shows that its log-det repulsion improves ACC and NMI by 1.2 and 1.9 percentage points over pairwise repulsion on CIFAR-10-LT with IR$=10$, while also improving tail-sample coverage. This evidence supports the resulting regularization design; the probabilistic formulation explains how its components jointly govern prototype capacity allocation.

We have added the comparison of repulsion formulations experiment in Appendix D.3.

Comments on the paper

1 comment

Dear Reviewers,

Thanks for your time and effort in reviewing this paper. The authors have responded to the comments by providing a rebuttal. Please read the rebuttal and indicate whether your concerns have been resolved. Thanks.

Regards from the AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.