acceptodds
Under review as a conference paper at ICLR 2027

Asynchronous CPU-GPU Density Peak Clustering via Granular-Ball Representation

Abstract

Density Peak Clustering (DPC) relies on extensive distance computations, leading to high time and memory costs when applied to large-scale data. Although existing granular-ball methods reduce the computational scale through coarse-grained representation, their scalability is still limited by the quality of granular-ball construction and the sequential execution of subsequent computations. To address these issues, we propose an incremental CPU-GPU parallel density peak clustering method based on granular-ball representation, termed ICG-DPC. First, the data space is partitioned into local regions, where granular balls are adaptively refined according to local density and structural dispersion. The balls are further split along the principal direction of the data distribution to better preserve complex spatial structures, thereby compressing the original data into a much smaller set of granular balls. Then, an asynchronous CPU-GPU collaborative framework is developed, in which the CPU dynamically performs granular-ball refinement and structural maintenance, while the GPU parallelizes local neighborhood construction, density estimation, and density-peak computation. A version-consistency mechanism is further employed to coordinate dynamic structural updates with parallel computation. Experimental results on large-scale synthetic and real-world datasets demonstrate that ICG-DPC maintains competitive clustering performance while providing favorable computational efficiency and scalability.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
granular ball, density peak clustering, incremental clustering, CPU-GPU parallel, heterogeneous computing

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

The paper proposes ICG-DPC for large-scale density peak clustering. It uses regional sampling and adaptive granular-ball splitting to reduce computation. The CPU updates the ball structure while the GPU computes densities and density peaks asynchronously. Boundary correction and global center selection produce the final clusters. Experiments on ten datasets report clustering quality and runtime, including datasets with over ten million samples.

Critical strengths

The paper addresses a practical problem: reducing the cost of clustering large datasets. Assigning structural updates to the CPU and numerical computation to the GPU is a reasonable design.

Critical weaknesses

The main contributions are not directly validated. There is no comparison between synchronous and asynchronous versions of the same algorithm, and no component ablations. Hardware, parameter settings, timing boundaries, and variability across runs are not reported. Comparisons with closely related methods, including GB-DP, AGB-DP, and UP-DPC, are also missing.

Important algorithm rules and evaluation details remain unclear. The paper does not fully specify center selection, label correction, assignment of unsampled points, or handling of outdated GPU results. Some reported scores also need clarification. For example, DPC achieves ACC = 0.396 on SUSY. If it produces two clusters and ACC uses optimal label matching, this score should be at least 0.5. The actual cluster counts and evaluation procedure are therefore essential.

Questions for authors

How are ACC and NMI calculated, and how many clusters does each method produce? Can you explain the SUSY result above?

What are the exact splitting, center-selection, and label-correction rules? How are unsampled points assigned to final granular balls?

How do you prevent outdated GPU results from affecting the final output? Is a final consistency check performed?

Reviewer_aCkROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This article proposes ICG-DPC, a large-scale density peak clustering method based on granular sphere representation, using an asynchronous CPU-GPU framework. The method consists of two core parts: (1) a granular sphere refinement strategy for structure perception, which jointly considers local density and structural dispersion, and uses PCA principal direction splitting to preserve anisotropic structures; (2) Asynchronous CPU-GPU execution framework, where the CPU is responsible for granularity ball refinement and spatial index maintenance, and the GPU is responsible for neighborhood construction, density estimation, and density peak calculation, achieving concurrency through producer consumer queues. The experiment compared with k-means, DPC, FCAG, LSC-R, LSC-K, USPEC, GB-USC, GB-SC and other methods on multiple synthetic and real large-scale datasets (up to 11 million samples), and reported ACC, NMI and running time.

Critical strengths
  1. The problem has practical significance. Density peak clustering faces secondary distance calculation and memory bottlenecks on large-scale data, and granularity ball representation and CPU-GPU heterogeneous computing are reasonable solutions.

  2. Asynchronous framework design has some novelty. Decouple structural dependency operations (granular sphere refinement, index maintenance) from data parallel operations (density estimation, peak calculation), and execute them concurrently through producer consumer queues with clear thinking.

  3. The experimental scale is relatively large. Tests were conducted on datasets such as TB_10M (10 million samples) and HIGGS (11 million samples), indicating that the method has a certain degree of scalability.

Critical weaknesses
  1. Limited method innovation and insufficient technical depth. There have been several works on the combination of granular spheres and DPC (such as Cheng et al. 2023, 2024); Zhang et al. 2025), GPU accelerated DPC has also been developed with UP-DPC (Ma et al. 2024), Liu et al. 2023, and others. The core contribution of this article is to combine the existing granular ball DPC with asynchronous CPU-GPU scheduling, which leans more towards engineering optimization and lacks new theoretical or algorithmic insights. The asynchronous framework itself is not a completely new concept, and the paper has not demonstrated its essential advantages over synchronous pipelines.

  2. Key method details are missing and have poor reproducibility. A large number of key formulas and mechanisms in the main text have been placed in the appendix or directly omitted:

  3. Insufficient experimental comparison and lack of comparison with the latest methods.

  4. Lack of theoretical analysis. The paper does not provide convergence proof, complexity analysis (time/space), error bounds, or approximate guarantees. For conferences like ICLR, if the methods are primarily focused on systems and experiments, at least a clear comparison of complexity should be provided.

  5. There are many issues with writing and symbols. The citation format is confusing, with multiple occurrences of "?" (such as "according to local density and the relative distance to higher density samples?") in the table layout, which affects reading.

Questions for authors
  1. On methodological novelty and the essential advantage of the asynchronous framework: The core contribution of this paper is the combination of existing granular-ball DPC with asynchronous CPU–GPU scheduling, but the asynchronous framework itself is not a new concept. Please clarify explicitly: compared with existing granular-ball DPC works (e.g., Cheng et al. 2023, 2024; Zhang et al. 2025) and GPU-accelerated DPC methods (e.g., UP-DPC, Liu et al. 2023), what new theoretical or algorithmic insights does this paper provide beyond engineering optimization? What is the essential advantage of the asynchronous framework over a synchronous pipeline, and can this advantage be demonstrated mechanistically or theoretically rather than only through experimental speedup?

  2. On key methodological details, version consistency, and reproducibility: The main text does not fully provide key details such as the specific formula for local density ,how GPU tasks are partitioned into regions, how N_min and N_max are set, how the k_n neighborhood is constructed, and how the kernel bandwidth is determined. In addition, it is unclear how “version consistency” is guaranteed, how conflict detection and rollback are handled when the CPU updates granular balls while the GPU may be reading an old version, and what the specific algorithm for cross-region correction is. Please supplement these key formulas and algorithmic steps, and explain how version consistency and result correctness are ensured under asynchronous updates. Moreover, please fix the discontinuous equation numbering (jumping directly from (6) to (8)) and undefined symbols .

  3. On the sufficiency of experimental comparisons and ablation analysis: The experiments lack direct comparisons with the latest GPU-accelerated DPC methods (e.g., UP-DPC and the GPU-DPC of Liu et al. 2023), and the baselines are relatively old, lacking large-scale clustering methods from 2023–2025. Please add systematic comparisons with the latest methods and include an ablation study against the original DPC under the same granular-ball representation, so as to separate the respective contributions of the “granular-ball representation” and the “asynchronous framework.” In addition, in Table 2, ICG-DPC has lower ACC than USPEC on MNIST and lower than LSC-R/LSC-K on SUSY; the authors only broadly attribute this to “metric- and dataset-dependent,” which lacks in-depth analysis. Please provide a more thorough analysis.

  4. **On hyper-parameter sensitivity, quantification of asynchronous speedup, and information loss in high-dimensional data: The method introduces multiple hyper-parameters, including the sampling ratio , the split-threshold control , the region sizes N_min,N_max, and the neighborhood size k_n, but the paper does not report their default values, selection rationale, or sensitivity experiments. Please supplement a hyper-parameter sensitivity analysis. Meanwhile, Table 3 shows that ICG-DPC is not the fastest on some datasets (e.g., 2.49 s on Letter versus only 1.29 s for USPEC; 6.81 s on MNIST, comparable to 4.61 s for LSC-R). The paper claims that the asynchronous framework improves hardware utilization, but it does not report CPU time, GPU time, communication overhead, or queue waiting time, nor does it compare with a synchronous version. Please quantitatively verify the substantial speedup brought by the asynchronous design. In addition, the authors acknowledge in the conclusion that regional sampling may lose fine-grained structural information in high-dimensional or highly complex data, and the ACC of ICG-DPC on MNIST is only 0.446, lower than the 0.698 of USPEC. Please explain whether there is an effective solution or mitigation mechanism.

  5. On theoretical analysis and writing standards: The paper lacks convergence proof, time/space complexity analysis, error bounds, or approximation guarantees; for work that is mainly system- and experiment-oriented, at least a clear complexity comparison should be provided. Please supplement the corresponding theoretical analysis or complexity discussion. At the same time, there are many writing and notation issues: citation formats are inconsistent and “?” appears in places, Equation (7) is missing, notation is inconsistent between the main text and the appendix , some sentences are grammatically awkward, and table formatting is disordered.

Reviewer_62b1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ICG-DPC, an asynchronous CPU‑GPU density peak clustering framework based on granular‑ball representation. The method introduces a structure‑aware granular‑ball refinement strategy that combines local density, structural dispersion, and PCA‑guided splitting to preserve complex spatial structures. It further adopts an asynchronous CPU‑GPU architecture equipped with version‑consistency mechanisms for concurrent computation, alongside cross‑region boundary correction to remedy partition‑induced artifacts, so as to boost computational efficiency.

Critical strengths

The integration of granularball compression with heterogeneous CPUGPU computation offers a sensible direction for alleviating the computational burden of densitypeak clustering. Decoupling structural refinement from parallel density computation is wellmotivated from a practical perspective and beneficial for handling largescale datasets.

Critical weaknesses

It is unclear to what extent each individual component contributes to the observed accuracy gains and runtime performance.

Questions for authors

Beyond assembling existing granularball refinement and parallelization techniques, what core methodological novelties distinguish ICGDPC from prior granularballbased densitypeak clustering approaches?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.