Clustering-driven Pseudo-label Generator (CPG)

• CPG mechanisms use clustering-derived structures to generate, rank, and refine pseudo-labels for unlabeled data across diverse tasks.
• They integrate strategies such as density estimation, prototype alignment, and consensus filtering to assess label reliability and set effective thresholds.
• Empirical studies show that embedding CPG modules into training cycles significantly enhances model supervision and performance in semi-supervised and unsupervised learning scenarios.

Clustering-driven Pseudo-label Generator (CPG) denotes a family of mechanisms in which clustering structure is used to create, filter, rank, or stabilize pseudo-labels for unlabeled data. In the literature, the term can refer to an explicit plug-and-play module, as in STAR-IOD, or to a broader architectural pattern embedded in semi-supervised learning, domain adaptation, re-identification, image clustering, continual learning, and label-free coreset selection (Zhang et al., 20 May 2026, Choi et al., 2019, Yan et al., 2022, Niu et al., 2021, He et al., 2021, Zheng et al., 2024). Across these settings, clustering is not restricted to assigning cluster IDs as labels; it may instead estimate reliability, align label spaces across time, define class-specific thresholds, or group semantically similar labels. One 2025 paper uses the acronym CPG differently, meaning “Controllable Pseudo-label Generation” rather than clustering-driven generation (Hou et al., 5 Oct 2025).

1. Terminology and scope

The terminology is not standardized. In STAR-IOD, the “Clustering-driven Pseudo-label Generator” is explicitly defined as a module that leverages K-Means clustering to dynamically identify class-specific thresholds for pseudo-labeling old classes during remote sensing incremental object detection (Zhang et al., 20 May 2026). In several adjacent lines of work, however, the same functional role appears without the exact name: density-ranked pseudo-label curricula in unsupervised domain adaptation, graph-corrected pseudo labels in person Re-ID, temporal pseudo-label consensus over generations, selective agreement filtering in deep clustering, and semantic prototype-based label generation in image clustering (Choi et al., 2019, Yan et al., 2022, Zhang et al., 2021, Mahon et al., 2021, Niu et al., 2021).

A precursor to this design logic appears in transformation-based unsupervised clustering, where pseudo parent-classes are randomly assigned and coupled to domain-specific transformations; ACOL and GAR then induce latent subclass specialization and produce a $k$-means-friendly representation (Kilinc et al., 2018). This suggests that CPG is better understood as a methodological family than as a single canonical algorithm.

The acronym itself is ambiguous. “Keep It on a Leash” uses CPG to denote “Controllable Pseudo-label Generation” for long-tailed semi-supervised learning, and its mechanism is based on two-view confidence-and-agreement filtering plus logit adjustment rather than clustering (Hou et al., 5 Oct 2025). By contrast, the supplied document for ProtoCon does not contain method content about ProtoCon at all, so it does not provide extractable evidence for a CPG interpretation of that method (Nassar et al., 2023).

2. Common structure of a CPG

Despite substantial variation, most CPG-style systems follow a recurrent structure. A representation is first computed for unlabeled data; clustering or a clustering-derived statistic is then used to propose pseudo-labels or pseudo-label utilities; a second mechanism converts that structure into supervision, filtering, or curriculum decisions; the resulting targets are injected into the training loop; and the improved model later regenerates more informative cluster structure. This suggests a general pattern of representation update $\rightarrow$ clustering-derived pseudo supervision $\rightarrow$ model update $\rightarrow$ renewed clustering-derived supervision (Choi et al., 2019, Zhang et al., 2021, Zia-ur-Rehman et al., 2024, Mahon et al., 2021).

The clustered object is not always the same. Some methods cluster feature embeddings; STAR-IOD clusters scalar teacher confidence scores; SemCo clusters label embeddings; RLCC and SLR cluster implicitly across time by constructing overlap-based mappings between consecutive clusterings; SPC aggregates an ensemble of clusterings through label alignment and agreement; ELFS uses deep clustering to create a surrogate label space for difficulty estimation (Zhang et al., 20 May 2026, Nassar et al., 2021, Zhang et al., 2021, Zia-ur-Rehman et al., 2024, Mahon et al., 2021, Zheng et al., 2024).

Representative method	What is clustered	Pseudo-label role
PCDA (Choi et al., 2019)	Target features within predicted class	Reliability ranking and curriculum
STAR-IOD (Zhang et al., 20 May 2026)	Class-wise teacher confidence scores	Class-specific thresholding of pseudo boxes
SPICE (Niu et al., 2021)	Prototype-centered feature neighborhoods	Semantic pseudo-label assignment and filtering
SPC (Mahon et al., 2021)	Ensemble clusterings of learned embeddings	Consensus labels and agreement mask
RLCC / SLR (Zhang et al., 2021, Zia-ur-Rehman et al., 2024)	Overlap between consecutive clusterings	Temporal projection and refined labels
SemCo (Nassar et al., 2021)	Label embeddings via DBSCAN	Grouped label-space pseudo-label refinement

A central implication is that “clustering-driven” does not necessarily mean “cluster ID equals final label.” In several systems, clustering is instead a latent control signal that determines whether a pseudo-label should be trusted, when it should be introduced, or how it should be converted into a soft target.

3. Principal algorithmic families

One major family uses clustering as a reliability scheduler. PCDA first obtains tentative pseudo-labels from a classifier, partitions target features by predicted category, computes pairwise distances $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$, defines local density $\rho_i$ by neighbor counting within a cutoff $e_c$, selects the highest-density point as class center, and applies $k$-means on distances to that center to split each predicted category into $P$ reliability groups; with $P=3$, these become the easy, moderate, and hard subsets $\rightarrow$0 that drive the curriculum (Choi et al., 2019). In this formulation, clustering does not generate class identities from scratch; it ranks classifier-generated pseudo-labels by geometric trustworthiness.

A second family uses semantic prototypes or cluster-friendly embeddings to generate pseudo labels directly. SPICE selects the top $\rightarrow$1 confident samples per cluster prediction, averages their frozen features into prototypes $\rightarrow$2, and assigns semantic pseudo-labels to nearby samples by cosine similarity to those prototypes; its later reliable pseudo-labeling stage retains sample $\rightarrow$3 only if the local consistency score $\rightarrow$4 exceeds a threshold $\rightarrow$5 (Niu et al., 2021). PL-CFE addresses the same problem upstream: it first learns a clustering-friendly embedding by minimizing an inter- to intra-class similarity ratio surrogate and then applies $\rightarrow$6-means with 500 clusters to obtain pseudo-labels for unsupervised meta-learning (Dong et al., 2022). SemCo operates in label space rather than instance space, clustering class embeddings with DBSCAN and thresholding group scores instead of per-class scores, so that visually similar classes can share pseudo-label confidence (Nassar et al., 2021).

A third family stabilizes pseudo-labels over time through cross-generation or cross-epoch consensus. RLCC computes a consensus matrix $\rightarrow$7, row-normalizes it into a transition matrix from old pseudo-label space to new pseudo-label space, propagates previous hard or soft pseudo labels by $\rightarrow$8, and ensembles them with current labels via $\rightarrow$9 (Zhang et al., 2021). SLR adopts the same IoU-overlap principle across consecutive epochs, forms $\rightarrow$0, and then applies HDBSCAN on the refined soft labels instead of taking an argmax (Zia-ur-Rehman et al., 2024). Both methods treat changing cluster vocabularies as a label-space alignment problem rather than as a defect to be ignored.

A fourth family uses graph correction or ensemble agreement to prune noisy pseudo labels. GLC starts from pseudo labels generated by any clustering backend, builds a $\rightarrow$1-NN graph using the joint similarity $\rightarrow$2, learns edge correctness with a 1-layer GCN and focal loss, prunes the graph by edge confidence and node connectivity, and recovers refined pseudo labels as connected components (Yan et al., 2022). SPC instead runs an ensemble of clustering pipelines, aligns their cluster IDs by the Hungarian algorithm, defines the consensus label as the mode across aligned assignments, and accepts a sample only when all ensemble members agree; selected samples receive cross-entropy supervision, while unselected samples remain under reconstruction loss (Mahon et al., 2021).

A fifth family uses clustering in score space rather than feature space. STAR-IOD collects teacher prediction scores $\rightarrow$3 for each old class, stores them in a class-specific memory bank $\rightarrow$4, applies 1D two-cluster K-Means to separate low- and high-confidence modes, sets the class-specific threshold to the lower bound of the high-confidence cluster $\rightarrow$5, and keeps old-class pseudo boxes satisfying $\rightarrow$6 before overlap filtering with $\rightarrow$7 (Zhang et al., 20 May 2026). Here clustering is a threshold estimator, not a semantic partitioner.

4. Relation to training objectives and system architectures

CPG mechanisms are typically embedded into larger training systems rather than used as isolated post-processors. In STAR-IOD, pseudo-labeled old-class boxes are merged with new-class annotations and injected into the detector’s training targets, thereby altering Hungarian assignment and the standard DETR-style loss $\rightarrow$8; the same pseudo labels also support STD because old-class prototypes are computed from pseudo-labeled instances in the current batch (Zhang et al., 20 May 2026). In PCDA, clustering-derived curricula control which subsets $\rightarrow$9, $\rightarrow$0, and $\rightarrow$1 participate in target training, and the Euclidean-based clustering loss $\rightarrow$2 is added to compact same-pseudo-label representations and separate different ones (Choi et al., 2019).

In re-identification, CPG-style refiners often become the supervision layer itself. RLCC replaces brittle hard cluster assignments with temporally ensembled soft targets in a SpCL framework, using prototype similarities and soft cross-entropy in the current pseudo-label space (Zhang et al., 2021). SLR uses its refined hard labels to drive classification, soft cross-branch supervision, and soft triplet terms, with the refinement beginning only from the second epoch because epoch 1 lacks a previous clustering for projection (Zia-ur-Rehman et al., 2024). MMT follows a related but distinct design: hard pseudo identities are produced offline by clustering every epoch, while online EMA teachers generate soft class and soft triplet pseudo labels for peer networks, yielding a mixed hard-and-soft supervision regime rather than a single refined label source (Ge et al., 2020).

Some systems place pseudo-label generation inside non-classification objectives. In slcGAN, a clustering network $\rightarrow$3 outputs soft real-image pseudo-label distributions $\rightarrow$4, the generator is conditioned on sampled one-hot pseudo labels, and the discriminator judges real and fake image-label pairs under a joint objective $\rightarrow$5 (Noroozi, 2020). The clustering network is trained on real images via adversarial pairing and augmentation consistency, while the generator is forced to produce fake images whose labels are recognizable to that clusterer. This is a generative interpretation of CPG in which pseudo-labels define useful conditional modes rather than downstream class targets.

The same pattern extends beyond direct model supervision. ELFS uses deep clustering via TEMI to assign hard pseudo-labels $\rightarrow$6, then trains a classifier on $\rightarrow$7 in order to compute proxy training-dynamics scores such as forgetting and AUM for coreset selection (Zheng et al., 2024). In that setting, the pseudo-label generator does not supervise the final model directly; it creates an auxiliary label space that makes otherwise supervised difficulty scores available in a label-free regime.

5. Applications and empirical performance

CPG mechanisms have been evaluated across markedly different tasks. In unsupervised domain adaptation, PCDA reports average accuracies of 88.3 on Office-31, 88.7 on imageCLEF-DA, and 65.8 on Office-Home when ECL is included, and its ablations show that progressively adding moderate and hard subsets improves over training on the easy subset alone (Choi et al., 2019). In person and object Re-ID, RLCC improves SpCL from 73.1 to 77.7 mAP on Market-1501 and from 19.1 to 27.9 mAP on MSMT17, while SLR improves ABMT from 67.8 to 79.1 mAP on Market-to-PersonX with ResNet50 and from 74.9 to 86.1 with IBN-ResNet50 (Zhang et al., 2021, Zia-ur-Rehman et al., 2024).

In detection, STAR-IOD reports gains of 1.7% and 2.1% mAP on DIOR-IOD and DOTA-IOD, respectively, and its module ablation shows that adding CPG alone raises $\rightarrow$8 from 1.5 to 30.4, indicating that pseudo-label recovery for missing old-class annotations is a dominant stabilization mechanism in remote sensing incremental detection (Zhang et al., 20 May 2026). In deep image clustering, SPICE reaches 0.926 ACC on CIFAR-10 and 0.938 ACC on STL10 in the whole-dataset setting, while SPC reaches 99.03% ACC and 97.04% NMI on MNIST through consensus-based selective pseudo-labeling (Niu et al., 2021, Mahon et al., 2021).

In unsupervised continual learning, global clustering serves as a simple CPG that allows standard supervised class-incremental methods to operate without labels after the first task. The crucial empirical finding is that the best feature extractor for clustering is the updated model from the previous step, $\rightarrow$9, and that fixed pseudo labels within an incremental step outperform iterative pseudo-label refreshing; on CIFAR-100 with step size 5, the one-shot method reaches 0.558 average ACC compared with 0.498, 0.523, and 0.513 for pseudo-label updates every 10, 20, and 30 epochs, respectively (He et al., 2021).

A further pattern is that pseudo-label utility is task-dependent. ELFS reports pseudo-label misclassification rates of 33.69% on CIFAR100 and 41.2% on ImageNet-1K with DINO features, yet those labels still support stronger label-free coreset selection than geometry-only alternatives (Zheng et al., 2024). This suggests that in some downstream tasks, the decisive property of a CPG is not exact semantic correctness but whether the induced pseudo-label space preserves enough structure to drive useful optimization signals.

6. Limitations, misconceptions, and unresolved issues

A common misconception is that a CPG always means “cluster IDs are directly used as labels.” The surveyed literature shows several counterexamples. PCDA uses clustering mainly to rank pseudo-label reliability rather than to define class identity (Choi et al., 2019). STAR-IOD clusters confidence scores in one dimension, not image embeddings (Zhang et al., 20 May 2026). SemCo clusters label embeddings rather than samples (Nassar et al., 2021). ELFS uses pseudo labels primarily to estimate training dynamics for coreset selection rather than to supervise the final model (Zheng et al., 2024). The term therefore denotes a broader design space in which clustering structure informs pseudo-label usage.

Another misconception is that better clustering metrics automatically imply better downstream behavior. RLCC, SLR, and MMT all emphasize that pseudo labels must remain stable under temporal drift; otherwise reclustering can amplify forgetting, confirmation bias, or label flipping (Zhang et al., 2021, Zia-ur-Rehman et al., 2024, Ge et al., 2020). SPC formalizes a related point theoretically: more accurate pseudo labels improve a clusterability measure $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$0, but the mechanism is selective use of reliable labels rather than indiscriminate use of all assignments (Mahon et al., 2021).

The main technical limitations recur across papers. Many methods depend strongly on the quality of the initial representation or teacher model; PCDA notes that poor early pseudo labels compromise within-class density estimates, STAR-IOD depends on the quality of teacher predictions, and GLC is explicitly less effective when clustering or graph construction is too weak early in training (Choi et al., 2019, Zhang et al., 20 May 2026, Yan et al., 2022). Several methods assume the number of clusters or classes is known: PL-CFE fixes $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$1, ELFS sets the number of clusters equal to the dataset class count $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$2, SPICE assumes known $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$3, and unsupervised continual learning sets cluster count equal to the known step size (Dong et al., 2022, Zheng et al., 2024, Niu et al., 2021, He et al., 2021). Hyperparameters such as $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$4, $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$5, $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$6, $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$7, $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$8, $\mathbf E_{ij}=\|f_i^t-f_j^t\|^2$9, or HDBSCAN min-cluster-size are also materially important in multiple systems (Choi et al., 2019, Zhang et al., 20 May 2026, Zia-ur-Rehman et al., 2024).

Open directions are implied rather than fully solved. Multi-epoch or multi-generation memory beyond adjacent steps remains limited in temporal refiners; SLR uses only epoch $\rho_i$0, and RLCC aligns only consecutive generations (Zia-ur-Rehman et al., 2024, Zhang et al., 2021). Sample-level uncertainty modeling is often rudimentary, especially when hard thresholds or hard label selections are used. The ambiguity of the term itself remains unresolved, because the same acronym can refer either to clustering-driven pseudo-labeling or to controllable pseudo-label generation without clustering (Hou et al., 5 Oct 2025). The literature therefore supports an operational definition: a CPG is best understood as any module in which clustering-derived structure governs the generation, acceptance, or refinement of pseudo-label supervision, rather than as a single standardized algorithm.