Domain Pseudo-label Mining & Latent Clustering

• Domain Pseudo-label Mining and Latent Clustering involves generating high-confidence surrogate labels from unlabeled data using iterative, confidence-aware clustering techniques.
• Graph- and ensemble-based approaches enhance label reliability by leveraging multiple views and consensus strategies to form robust latent structures.
• Optimal transport and hierarchical clustering methods reduce noise and mitigate domain shift, leading to measurable gains in domain adaptation and cross-domain representation.

Domain pseudo-label mining and latent clustering are foundational techniques underpinning recent advances in unsupervised and semi-supervised domain adaptation, domain generalization, and cross-domain representation learning. These approaches focus on the generation and refinement of surrogate labels (“pseudo-labels”) for unlabeled target or auxiliary domains, and the discovery of latent structure—often viewed as clusters or “pseudo-domains”—within high-dimensional feature spaces. Modern methods leverage these constructs to mitigate domain shift, extract robust representations, and enhance transferability across diverse modalities and tasks.

1. Principles of Domain Pseudo-Label Mining

Domain pseudo-label mining refers to the automatic assignment of discrete labels (or soft surrogates) to unlabeled target samples, typically by clustering in the learned latent space or via model-ensemble consensus. These pseudo-labels serve as training targets, enabling supervised objectives in the absence of ground-truth annotations and facilitating conditional alignment across domains.

Key principles include:

• Confidence-aware selection: As noisy pseudo-labels can harm learning, several methods focus on mining only high-confidence labels, e.g., selective training on points with multi-view agreement (Mahon et al., 2021), or using statistical assessments to guard against false-merges (Li et al., 2023).
• Iterative refinement: Most pipelines alternate between label assignment and feature learning, interleaving clustering or optimal transport-based assignment steps with representation updates (Zheng et al., 2021, Zhang et al., 2022).
• Domain and cluster correspondence: Pseudo-labels may be constructed to reflect either semantic categories or domain structure. In the latter case, “latent domain discovery” seeks to partition an unlabeled or mixed-domain set into coherent sub-domains without access to external domain tags (Thomas et al., 9 Mar 2025, Chen et al., 25 Mar 2026).

2. Graph- and Ensemble-Based Pseudo-Labeling

Graph-based and ensemble-driven mechanisms enhance the reliability and consistency of pseudo-label assignments, particularly in high-noise or large-scale settings.

• Progressive Sub-Graph Clustering: The PGMVG algorithm (Li et al., 2023) constructs multiple k-nearest-neighbor graphs over target embeddings extracted from a diverse set of pre-trained models. Only edges present across all models are retained via intersection, yielding a robust union graph. Pseudo-labels are defined by connected components at a small initial $k_0$, then progressively refined: $k$ is increased iteratively, adding new edges and merging components subject to a double-Gaussian test on model-level pairwise similarities. This process balances sensitivity (label expansion) and specificity (error rejection), with statistical safeguards against catastrophic merges.
• Selective Consensus: The Selective Pseudo-label Clustering (SPC) scheme (Mahon et al., 2021) exploits an ensemble of $K$ clustering pipelines: only samples with unanimous agreement are used for supervised updates, while the remainder contribute only via reconstruction losses. Theoretical analysis demonstrates that this selective mask both reduces noise and accelerates formation of k-means-friendly latent embeddings.

3. Optimal Transport and Assignment-Based Label Transfer

Optimal transport and assignment-based strategies enable both the efficient reallocation of pseudo-labels and the correction of accumulated noise.

• Group-aware Label Transfer (GLT): GLT (Zheng et al., 2021) models pseudo-label allocation as a regularized optimal transport problem, subject to balanced marginals that promote cluster size uniformity and prevent collapse. Critically, GLT applies this process in parallel across multiple “attribute groupings” (multi-scale clusterings of the latent space), constraining label transfer and enhancing robustness to noisy assignments.
• Class-aware UDA (CA-UDA): CA-UDA (Zhang et al., 2022) clusters target features, performs Hungarian-matched optimal assignment to align source and target class centroids, and then iteratively refines pseudo-labels with a target-only auxiliary classifier using self-paced “easy-to-hard” learning. This mitigates source-domain bias and yields tightly clustered, semantically aligned representations.

4. Stable Latent Clustering and Domain Discovery

Flat clustering in non-stationary or noisy feature spaces is often unstable. Hierarchical and representative-based schemes address this limitation.

• Granular Ball Clustering: In domain-general crowd counting (Chen et al., 25 Mar 2026), hierarchical “granular ball” construction replaces direct k-means on all features. Local, compact “balls” are recursively split only if intra-ball compactness improves. Final pseudo-domains result from clustering a small set of ball centers, which is both robust to feature drift and stabilizes pseudo-domain assignments epoch-to-epoch.
• Diffusion Latent Space Clustering: Diffusion models offer exceptionally rich domain-variant latent representations. Extracting latent codes from a frozen diffusion backbone, followed by k-means clustering, enables unsupervised pseudo-domain assignment (Thomas et al., 9 Mar 2025). These pseudo-domains serve as auxiliary signals: a ridge-mapped summary of each centroid is concatenated to the main backbone features, enhancing domain robustness for downstream classifiers.

5. Pseudo-Label Noise Mitigation and Refinement

Given the prevalence of label noise, pseudo-label refinement methodologies are central to effective domain adaptation.

• Double-Gaussian Statistical Assessment: During iterative subgraph merging (Li et al., 2023), merging candidate clusters is subjected to multi-model double-Gaussian fitting over the distribution of pairwise similarities. Merging is performed only if the resulting distributions are unimodal or sufficiently overlapping at high similarity, sharply reducing the rate of erroneous merges.
• Intra-Class Similarity and Spanning Trees: Pseudo-labels are refined by thresholding pairwise intra-class similarities, removing low-confidence nodes, and then discovering high-confidence cores via minimum spanning trees in the graph of target features (Wang et al., 2022). Low-confidence instances are re-labeled by strong classifiers trained on these cores, iteratively boosting overall label quality.
• Tri-branch and Mutual Mean-Teaching: In semantic segmentation (Xu et al., 2022) and person re-ID (Ge et al., 2020), ensembles or dual-teacher frameworks train each model by pseudo-labels derived from the agreement of distinct branches or moving-averaged teacher nets. This triangulation reduces confirmation bias and accumulates more robust supervisory signals.

6. Applications and Empirical Impact

The methodologies detailed above are realized across diverse domains including speaker verification, person and vehicle re-identification, semantic segmentation, crowd counting, and domain-generalized image classification.

Empirical findings demonstrate:

• Significant performance gains over single-model or naive k-means baselines using multi-model, statistically robust, or confidence-selective pseudo-label mining (Li et al., 2023, Mahon et al., 2021, Zheng et al., 2021).
• Tighter latent cluster formation and improved intra-class compactness/inter-class separation, both visually (e.g., t-SNE, UMAP) and quantitatively, after application of these advanced clustering/regulatory routines (Zhang et al., 2022, Zheng et al., 2021, Chen et al., 25 Mar 2026).
• Gains in end-task accuracy/mAP exceeding prior state-of-the-art approaches in benchmark domain adaptation and generalization scenarios (e.g., +4.3% accuracy on DomainNet (Li et al., 2021), +4% absolute on PACS (Thomas et al., 9 Mar 2025), or +3–13% mAP in re-ID (Ge et al., 2020)).

A summary of distinctive approaches is presented below:

Method	Pseudo-Labeling Strategy	Latent Clustering Procedure
PGMVG (Li et al., 2023)	Multi-model voting, double-Gaussian merge-test	Iterative k-NN subgraph, component merging
GLT (Zheng et al., 2021)	OT-based assignment, group-wise constraining	Multi-scale groupings (K-means/DBSCAN + Sinkhorn)
CA-UDA (Zhang et al., 2022)	Hungarian assignment, self-paced refinement	K-means centroids, feature space alignment
Granular Ball (Chen et al., 25 Mar 2026)	Hierarchical ball partition, compactness filtering	Cluster ball centers, representative-centric
GUIDE (Thomas et al., 9 Mar 2025)	K-means++ in diffusion latent space	RBF mapping of centroids, domain augmentation

7. Limitations, Open Problems, and Directions

Challenges persist in domain pseudo-label mining and latent clustering:

• Stability and scalability: Many procedures require careful tuning of hyperparameters (e.g., initial $k_0$, threshold values, number of groups) and some, such as granular-ball clustering or repeated multi-model fitting, introduce additional computational overhead (Li et al., 2023, Chen et al., 25 Mar 2026).
• Noisy or dynamic feature distributions: Pseudo-label stability can be compromised by representation drift, outliers, or rapid changes in the feature manifold. Hierarchical or ensemble methods partly address these but are not universally robust.
• Domain overlap and label mismatch: Methods assuming shared label sets or fully covering pseudo-domain partitions may degrade when classes or domains are only partially overlapping (Rozner et al., 2023, Thomas et al., 9 Mar 2025).
• Integration with model design: Some frameworks treat clustering and representation learning as loosely coupled, which can slow convergence. Joint or fully differentiable approaches (e.g., via contrastive learning or differentiable clustering) remain active research directions.

A plausible implication is that future frameworks will increasingly integrate robust, sample-efficient, and end-to-end differentiable pseudo-label mining schemes with explicit uncertainty modeling and scalable manifold discovery, tailored for extreme domain heterogeneity and non-stationarity.