Prototype-Guided Clustering Methods

• Prototype-guided clustering is a family of methods that use representative anchors, defined as geometric, probabilistic, or functional prototypes, to drive and regularize the clustering process.
• These approaches employ alternating assignment-update schemes, leveraging EM-style optimization and contrastive losses to ensure robust cluster formation and separation.
• Empirical studies show that prototype-guided clustering improves performance metrics like NMI and ARI across applications including deep image segmentation, transfer learning, and federated settings.

Prototype-guided clustering encompasses a diverse family of methodologies in which cluster representatives, or “prototypes,” are used to steer, regularize, or explicitly determine the clustering process. These prototypes act as geometric, probabilistic, or functional anchors, often with properties or update rules distinct from instance assignments. The prototype paradigm appears in classical algorithms (e.g., k-means, FCM), but modern advances extend and generalize the concept to deep embedding spaces, structured output settings, federated and transfer learning, and self-supervised regimes. The following sections detail the mathematical formulations, algorithmic structures, empirical properties, and theoretical results from contemporary research in prototype-guided clustering.

1. Mathematical Foundations and Prototype Definitions

Prototype-guided clustering begins with the abstraction of a prototype: a vector (or set of vectors) that represents, anchors, or generates a cluster. In classical k-means, prototypes are means in $\mathbb{R}^d$, updated via arithmetic averaging. Fuzzy c-means (FCM) generalizes the notion to soft assignment with fuzzy partitions and includes source-domain prototypes for transfer clustering (Deng et al., 2014).

Recent deep clustering methods formalize prototypes as explicit learnable parameters, as centers in a latent space (e.g., $\{\mu_k\}_{k=1}^K$), or as output of a differentiable function of features (e.g., via a Dual Competitive Layer in DDCL (Cirrincione, 2 Apr 2026)). In probabilistic frameworks, prototypes may be means of Gaussian or vMF components, and clustering is achieved via assigning points to their nearest prototype under a specified similarity measure (Euclidean, cosine, or domain-adapted).

Prototypes can be hard (fixed centroids) or soft (weighted averages with assignment probabilities), static or dynamically updated, global or distributed (per-client in federated scenarios (Zou et al., 10 Mar 2026)). Advanced methods leverage multi-prototype representations per cluster or class, dynamically regulate the number of prototypes, or enable prototype transfer across domains or tasks (Qu et al., 10 Feb 2025, Deng et al., 2014).

2. Prototype Computation and Update Mechanisms

Prototype computation varies with methodology, but most share the alternating assignment–update schema. In classical settings:

• k-means: Assign each point to nearest prototype, then update prototypes to the mean of assigned points.
• Fuzzy c-means (FCM): Assign soft memberships $u_{ij}$; prototype $v_i$ is updated via $v_i = \frac{\sum_j u_{ij}^m x_j}{\sum_j u_{ij}^m}$ (Deng et al., 2014).

In transfer settings, additional regularization terms pull prototypes toward prior (source) prototypes: $$v_i = \frac{\sum_j u_{ij}^m x_j + \lambda_1 \sum_j u_{ij}^m \tilde{v}_i + \lambda_2 \tilde{v}_i}{(1+\lambda_1) \sum_j u_{ij}^m + \lambda_2}$$ (Deng et al., 2014).

In deep/probabilistic clustering:

• CPCC (Center-Oriented Prototype Contrastive Clustering) computes soft assignment $q_{ik}$ via a Student's t distribution over distances to k-means centers $\mu_k$, then builds soft prototypes by weighted sums $$p_k = \frac{\sum_i w_{ik} \mathcal{O}(x_i^t)}{\|\cdot\|_2}$$, with weights $w_{ik} \propto q_{ik}^2$ (Dong et al., 21 Aug 2025).
• ProPos alternates between spherical k-means cluster assignment (E-step) and prototype updates as normalized means of batch-embedded samples, combining alignment, prototype scattering, and positive sampling alignment (Huang et al., 2021).
• DDCL eliminates external clustering: the Dual Competitive Layer directly computes prototypes as outputs $\{\mu_k\}_{k=1}^K$0, permitting end-to-end differentiation through clustering (Cirrincione, 2 Apr 2026).
• Multi-prototype regimes (e.g., federated knowledge distillation in MP-FedKD (Zou et al., 10 Mar 2026)) use hierarchical clustering (CHAC) per class to produce multiple local prototypes and aggregate them in federated averaging.

Nonparametric or adaptive approaches may dynamically adjust the number of prototypes per cluster or class, e.g., BRPG's adaptive prototype optimization based on feature dispersion (Dong et al., 2023).

3. Prototype-based Clustering Objectives and Losses

Prototype-guided approaches define clustering objectives that explicitly depend on prototype–data relationships:

• Classical summaries: Minimize sum of squared distances to prototypes (k-means, FCM), or fuzzy-weighted analogs in FCM/TFSC (Deng et al., 2014).
• Regularized transfer: Include penalty terms for discrepancy between target and source-domain prototypes (parameter transfer) and for target memberships referencing source prototypes (label transfer) (Deng et al., 2014).
• Contrastive learning:
  • CPCC and ProPos introduce prototype-level InfoNCE losses, treating each prototype as an anchor and maximizing mutual information or alignment between corresponding prototype pairs while spreading prototypes apart, e.g., $\{\mu_k\}_{k=1}^K$1 in CPCC (Dong et al., 21 Aug 2025).
  • PSA (ProPos) aligns embeddings of augmented neighbors, strengthening cluster compactness (Huang et al., 2021).
• Variance-based separation:
  • DDCL decomposes the soft quantization loss into a least squares reconstruction and a non-negative variance term $\{\mu_k\}_{k=1}^K$2, whose gradient acts as a repulsive force, inherently preventing prototype collapse and guaranteeing Lyapunov stability for the frozen encoder system (Cirrincione, 2 Apr 2026).
• Federated and compositional settings:
  • LEMGP (MP-FedKD) loss combines an attraction term (between local embeddings and global prototypes) and a log-sum-exp repulsion from non-matching prototypes (Zou et al., 10 Mar 2026).
  • ClusPro uses contrastive and decorrelation losses based on assignment to within-primitive clustering discovered prototypes (Qu et al., 10 Feb 2025).

In segmentation and structure-aware modalities, prototypes may be constructed separately for high- and low-confidence features (boundary refinement) (Dong et al., 2023) or by superpixel-aware aggregation (Li et al., 2021).

4. End-to-End and EM-style Optimization Frameworks

The interplay of prototype updates and cluster assignments is realized through various algorithmic frameworks:

• Expectation–Maximization (EM): Many deep prototype-guided clustering methods alternate between E-steps (pseudo-label or probabilistic assignment via clustering) and M-steps (prototype construction, parameter update):
  • ProPos: spherical k-means E-step, then prototype alignments and scattering in M-step (Huang et al., 2021).
  • DigPro: EM alternates between full-dataset spherical k-means (E-step) and batch-wise prototype updates plus group-wise and prototype contrastive losses (M-step), updating the target network by EMA (Zhang et al., 2024).
  • BRPG: samples high- and low-confidence features, then runs per-class k-means in a sampled feature memory during specified epochs, updating prototypes dynamically (Dong et al., 2023).
• End-to-end differentiable clustering: DDCL directly incorporates prototype generation and assignment as a differentiable network layer, enabling global optimization without explicit clustering iterations (Cirrincione, 2 Apr 2026).
• Federated protocols: Multi-prototype clustering is executed locally (via CHAC), with per-class prototypes and updates, then aggregated globally, maintaining communication efficiency and non-IID robustness (Zou et al., 10 Mar 2026).

5. Regularization, Prototype Diversity, and Boundary Handling

Modern prototype-guided clustering schemes incorporate multiple mechanisms to enhance discrimination, mitigate prototype drift, and sharpen class boundaries:

• Prototype separation: Explicit uniformity terms or prototype–prototype repulsion penalties ensure that prototypes scatter across latent space (e.g., PSL in ProPos (Huang et al., 2021), $\{\mu_k\}_{k=1}^K$3 in DDCL (Cirrincione, 2 Apr 2026), $\{\mu_k\}_{k=1}^K$4 in ProtoGCD (Ma et al., 2 Apr 2025)).
• Boundary-aware prototypes: Partitioning features by confidence into high- and low-confidence memory banks enables dedicated boundary prototypes (BRPG (Dong et al., 2023)).
• Adaptive prototype count: Quantifying feature dispersion, e.g., by within-class cosine similarity or $\{\mu_k\}_{k=1}^K$5 norm, guides augmentation in prototype number per class or per cluster (Dong et al., 2023).
• Prototype drift minimization: Weighting contributions in prototype computation (e.g., $\{\mu_k\}_{k=1}^K$6 in CPCC) ensures that high-confidence, centrally-located embeddings dominate average, reducing the impact of boundary or noisy instances (Dong et al., 21 Aug 2025).
• Diversity in federated settings: Multi-prototype aggregation and prototype alignment (PA) mitigate the information loss from single-prototype collapse under data heterogeneity (Zou et al., 10 Mar 2026).

6. Empirical Performance and Application Domains

Prototype-guided clustering consistently demonstrates improvements in clustering quality and representation learning across a range of benchmarks and settings:

• Deep embedding clustering: ProPos, CPCC, DigPro, and DDCL all report superior normalized mutual information (NMI), adjusted Rand index (ARI), and accuracy (ACC) relative to non-prototype or instance-level contrastive methods, on datasets including CIFAR-10, CIFAR-20, STL-10, ImageNet-10/Dogs/Tiny-ImageNet (Huang et al., 2021, Dong et al., 21 Aug 2025, Zhang et al., 2024, Cirrincione, 2 Apr 2026).
• Federated clustering: MP-FedKD outperforms established baselines (FedProto, FedProx, MOON, etc.) in non-IID settings due to more accurate, robust class representation and convergence (Zou et al., 10 Mar 2026).
• Semantic segmentation: Boundary-refined and superpixel-guided prototype schemes (BRPG, ASGNet) yield consistent $\{\mu_k\}_{k=1}^K$7–$\{\mu_k\}_{k=1}^K$8 point mIoU gains on PASCAL VOC, Cityscapes, and COCO, especially for few-label or semi-supervised splits (Dong et al., 2023, Li et al., 2021).
• Transfer and domain adaptation: Prototype-oriented clustering and fuzzy transfer clustering show robust improvements in negative transfer mitigation and clustering quality on 20-Newsgroup, synthetic, and multidomain image benchmarks (Tanwisuth et al., 2023, Deng et al., 2014).
• Generalized category/class discovery: PNP, ProtoGCD, and related methods adapt prototype numbers dynamically and deliver strong improvements (up to $\{\mu_k\}_{k=1}^K$9 accuracy, $u_{ij}$0 clustering efficiency gain) on fine-grained and large-category discovery tasks (Wang et al., 2024, Ma et al., 2 Apr 2025).

Common trends include accelerated convergence, improved class separation, and greater robustness to cluster misspecification, non-IID, and boundary ambiguity. Prototype-based contrastive losses frequently outpace equivalent instance-contrastive or non-contrastive objectives. Adapting prototype count and leveraging potential/learnable prototypes prove critical in discovery and open-world settings.

7. Recent Theoretical Advances

A number of prototype-guided frameworks provide theoretical guarantees and analytic insight:

• Variance-driven separation: DDCL introduces an exact loss decomposition where the variance term’s gradient is an automatic “push-apart” force that maintains prototype diversity without explicit auxiliary objectives, yielding global Lyapunov stability for the frozen-encoder system and a provable negative feedback cycle regulating prototype dispersion (Cirrincione, 2 Apr 2026).
• Transfer clustering convergence: Transfer fuzzy c-means and subspace clustering exhibit guaranteed monotonic cost reduction and local convergence by Zangwill’s theorem (Deng et al., 2014).
• Federated and compositional properties: In federated clustering, hierarchical clustering (CHAC) guarantees robust within-class mode capture and stable, low-variance aggregation (backed by empirical convergence evaluations) (Zou et al., 10 Mar 2026).
• EM-optimization guarantees: For methods based on EM alternation (ProPos, DigPro), each M-step increases the likelihood or tightens its lower bound; spherical k-means E-steps ensure representation—cluster assignments are regularly updated for global compactness and separation (Huang et al., 2021, Zhang et al., 2024).

No single framework resolves all theoretical gaps, especially for global convergence with co-trained deep backbones. However, prototype-guided clustering constitutes a principal approach for effective unsupervised and semi-supervised representation learning, robust clustering under domain, data, and modality shifts, and interpretable, modular assignment-based structure discovery in both classical and deep learning settings.