Online Clustering Framework (OCF)

• OCF is a family of online clustering paradigms designed to update cluster assignments incrementally as new data arrives under constraints such as latency and limited memory.
• These methods often enforce irreversible merge operations, balancing statistical fidelity with operational feasibility in dynamic, streaming environments.
• Recent approaches expand OCF using probabilistic, neural, and task-specific techniques, addressing challenges in areas like speaker diarization and adaptive prototype learning.

Online Clustering Framework (OCF) denotes a class of clustering formulations in which data, relations, or decisions are revealed incrementally and the clustering must be maintained under streaming, partial-information, latency, or memory constraints. Across the literature, the term encompasses several non-equivalent but structurally related settings: sequential graph clustering with irreversible agglomeration, online Euclidean clustering with a changing number of centers, mini-batch deep clustering, low-latency speaker diarization, fixed-confidence bandit identification of cluster structure, neuromorphic clustering on spike streams, and bounded-memory interpretation of long seafloor image sequences (Mathieu et al., 2010, Chrobak et al., 2014, Li et al., 2016, Chen et al., 2022, Smith, 2023, Yan et al., 2024, Chandran et al., 20 Jan 2025, Liang et al., 8 Sep 2025). This suggests that OCF is best understood as a family of online clustering paradigms rather than a single standardized algorithm.

1. Scope, defining constraints, and recurring design pattern

The common feature of OCF formulations is that clustering is no longer a one-shot partition of a fixed dataset. Instead, the framework receives information sequentially and must update a maintained clustering state. In graph-based formulations, new vertices arrive and reveal relations only to previous vertices; in streaming vector formulations, a new point $x_t$ arrives and the learner must incur clustering loss immediately; in task-specific systems such as speaker diarization and seafloor imagery, latency and bounded memory become explicit architectural constraints (Mathieu et al., 2010, Li et al., 2016, Chen et al., 2022, Liang et al., 8 Sep 2025).

A second recurring property is restricted editability. Several classical online formulations allow cluster creation and cluster merging but disallow later splitting of already merged items. That irrevocability is central in online correlation clustering and online clique clustering, where a locally appealing merge can permanently preclude a globally better structure (Mathieu et al., 2010, Chrobak et al., 2014). By contrast, newer task-driven systems sometimes reintroduce controlled refinement through split mechanisms, beam search, or posterior resampling, but only under bounded computation or summary-memory budgets (Li et al., 2016, Chen et al., 2022, Liang et al., 8 Sep 2025).

A third common element is an explicit tension between statistical fidelity and operational feasibility. Different OCFs resolve that tension through competitive analysis, regret bounds, stochastic approximation, beam truncation, representative memory, or mini-batch self-labeling. The framework is therefore defined as much by its update constraints as by its clustering objective.

Setting	Online unit	Characteristic mechanism
"Online Correlation Clustering" (Mathieu et al., 2010)	One vertex at a time	Create singleton, merge clusters, never split
"Online Clique Clustering" (Chrobak et al., 2014)	One vertex at a time	Irrevocable clique-preserving merges
"A Quasi-Bayesian Perspective to Online Clustering" (Li et al., 2016)	One point $x_t$ at a time	Gibbs quasi-posterior over variable cluster count
"Interrelate Training and Searching: A Unified Online Clustering Framework for Speaker Diarization" (Chen et al., 2022)	Incremental speaker embeddings	CGRT plus TBSC with bounded latency
"Deep Online Probability Aggregation Clustering" (Yan et al., 2024)	Mini-batch features	Online probability aggregation in probability space
"Online Clustering of Seafloor Imagery for Interpretation during Long-Term AUV Operations" (Liang et al., 8 Sep 2025)	Image batches from a stream	DPGMM, merge/split, fixed representative memory

2. Classical online formulations: graph agglomeration and objective asymmetry

A canonical OCF appears in online correlation clustering. Vertices $v_1,v_2,\dots,v_n$ arrive sequentially; when $v_t$ arrives, the signs of all edges between $v_t$ and $\{v_1,\dots,v_{t-1}\}$ are revealed. The algorithm may create a new singleton cluster for $v_t$, place it into an existing cluster, and merge existing clusters, but it may not split an already formed cluster. The two standard objectives are

$$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$

for MinDisAgree, and

$$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$

for MaxAgree. The natural Greedy algorithm repeatedly merges any pair of clusters with positive gain. For MaxAgree, Greedy is $0.5$-competitive; no deterministic online algorithm can be better than $x_t$0-competitive and no randomized online algorithm can be better than $x_t$1-competitive; a randomized mixture of Greedy and $x_t$2 attains $x_t$3. For MinDisAgree, Greedy is $x_t$4-competitive, and the online problem is only $x_t$5-competitive (Mathieu et al., 2010).

Online clique clustering tightens the structural constraint by requiring every cluster to be a clique. The online model again reveals one vertex at a time, does not know $x_t$6 in advance, and permits only clique-preserving merges after the new vertex first appears as a singleton. For maximization, the objective is

$x_t$7

while for minimization the cost is

$x_t$8

The paper shows that a natural greedy strategy is linear-competitive for maximization: its competitive ratio is at least $x_t$9. The main positive result is a doubling-technique strategy with asymptotic competitive ratio at most $v_1,v_2,\dots,v_n$0 and absolute competitive ratio at most $v_1,v_2,\dots,v_n$1; the paper also proves that no deterministic strategy can have an asymptotic competitive ratio better than $v_1,v_2,\dots,v_n$2. For the minimization objective, the deterministic competitive ratio is $v_1,v_2,\dots,v_n$3, and Greedy achieves absolute competitive ratio $v_1,v_2,\dots,v_n$4 (Chrobak et al., 2014).

These graph-based OCFs establish a foundational point: the online constraint does not simply weaken offline clustering by a constant factor. It can qualitatively change what is achievable, and the direction of the change depends strongly on the objective. Offline-equivalent viewpoints such as maximizing agreements versus minimizing disagreements cease to be equivalent once the framework imposes partial revelation and irreversible merges.

3. Probabilistic, Bayesian, and prototype-based adaptive frameworks

A distinct OCF lineage replaces irrevocable agglomeration by online probabilistic inference over clustering structures. In the quasi-Bayesian framework, a clustering is represented by centers

$v_1,v_2,\dots,v_n$5

and the instantaneous loss is the squared distance to the nearest center,

$v_1,v_2,\dots,v_n$6

The parameter space is the union $v_1,v_2,\dots,v_n$7, and a prior mixes over unknown cluster count $v_1,v_2,\dots,v_n$8 via

$v_1,v_2,\dots,v_n$9

At each time, the method samples from a Gibbs quasi-posterior

$v_t$0

with a variance-corrected energy recursion. The framework adapts the number of clusters over time, achieves regret of order $v_t$1, admits an adaptive schedule $v_t$2, and is supported by minimax lower bounds of the same rate. Its practical sampler, PACBO, is an RJMCMC approximation whose Markov chain is proved $v_t$3-irreducible, aperiodic, and Harris recurrent (Li et al., 2016).

Online Deterministic Annealing recasts OCF as a temperature-controlled sequence of stochastic approximation problems over prototypes $v_t$4. The central objective is the free energy

$v_t$5

with Gibbs assignment rule

$v_t$6

At high temperature, assignments are soft and effective prototypes collapse; as $v_t$7 decreases, bifurcation occurs when the Hessian of $v_t$8 loses positive definiteness. The framework implements growth by perturbing prototype pairs $v_t$9 and $v_t$0, retaining them when they separate. For Bregman divergences, prototype updates reduce to centroid-like recursions through sufficient statistics $v_t$1 and $v_t$2, yielding online estimates $v_t$3 under standard stochastic approximation conditions $v_t$4 and $v_t$5 (Mavridis et al., 2021).

Another adaptive OCF removes any need to specify the number of clusters in advance by learning Gaussian functions online and clustering the functions rather than the raw points. The $v_t$6-th Gaussian is

$v_t$7

and for equal widths its center update is

$v_t$8

Pairwise co-activation statistics accumulate in $v_t$9 through

$\{v_1,\dots,v_{t-1}\}$0

and cluster membership is then induced by thresholding the normalized correlation graph with $\{v_1,\dots,v_{t-1}\}$1. The resulting connected components allow arbitrary shaped clusters and avoid fixing cluster count a priori; the paper reports acceptable clustering on several toy datasets and notes that $\{v_1,\dots,v_{t-1}\}$2 produced the largest sets of acceptable thresholds $\{v_1,\dots,v_{t-1}\}$3 (Eidheim, 2023).

4. Neural and deep OCF architectures

Neuromorphic OCF implements clustering through active dendritic computation rather than explicit centroids or graph merges. A segment computes a dot product $\{v_1,\dots,v_{t-1}\}$4 and, with threshold $\{v_1,\dots,v_{t-1}\}$5, outputs the thresholded potential. A dendrite contains multiple parallel segments, applies winner-take-all inhibition, and emits a 1-hot integer vector whose winning index is the cluster identifier. For bit vectors of constant Hamming weight $\{v_1,\dots,v_{t-1}\}$6, the framework uses sum of absolute differences and derives that minimizing SAD is equivalent to maximizing a dot product with the centroid approximation. Online learning is governed by Spike Dependent Plasticity with mutually exclusive Capture, Backoff, and Search updates. The same substrate supports spike sorting and supervised online classification via multiple dendrites and a voting mechanism. In the MNIST experiment, final error on the 10K test inputs is about $\{v_1,\dots,v_{t-1}\}$7, and after adaptation it can reach as low as $\{v_1,\dots,v_{t-1}\}$8 in blocks; increasing the segment count from 8 to 24 lowers error from around $\{v_1,\dots,v_{t-1}\}$9 to around $v_t$0 (Smith, 2023).

Deep PAC formulates OCF in probability space rather than center space. Its centerless clustering objective is

$v_t$1

and the core iterative membership update is

$v_t$2

The online module, OPA, applies this idea within a mini-batch by computing a score matrix $v_t$3, sharpening it into target pseudo-labels $v_t$4, and training the network through $v_t$5. The framework thereby interleaves clustering and representation learning without repeated full-dataset K-means or FCM passes, and the paper emphasizes that OPA does not impose a cluster-size balancing constraint. This version of online clustering is therefore not irreversible agglomeration but continuous, centerless self-labeling over streamed mini-batches (Yan et al., 2024).

Taken together, these neural frameworks broaden the meaning of OCF. “Online” may refer to per-cycle local synaptic adaptation, as in dendritic clustering, or to batchwise pseudo-label aggregation within end-to-end deep optimization. In both cases, inference and learning are interleaved rather than separated into a fixed offline clustering stage followed by model training.

5. Task-specific OCF systems: diarization, bandits, and persistent exploration

In online speaker diarization, OCF is explicitly defined as a unified interaction between embedding learning and online clustering. The framework consists of Clustering-Guided Recurrent Training (CGRT) and Truncated Beam Searching Clustering (TBSC). CGRT clusters intermediate embeddings during training, matches the resulting clusters to ground truth through a perfect matching with weight

$v_t$6

and derives metric thresholds

$v_t$7

TBSC then performs beam search over online cluster assignments, using $v_t$8 and $v_t$9 to penalize scores for existing and new clusters, with $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$0 and $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$1. On AISHELL-4, the reported system achieves $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$2 DER with collar $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$3 at $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$4s latency, compared with $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$5 for offline agglomerative hierarchical clustering (Chen et al., 2022).

Bandit OCF studies pure-exploration clustering when only one arm can be sampled per round. Each arm produces i.i.d. Gaussian samples

$$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$6

and the target clustering is defined by applying SLINK to the unknown means under the separability condition $$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$7. The main algorithm, ATBOC, uses average tracking to approximate the optimal allocation

$$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$8

and stops according to the generalized likelihood ratio statistic

$$cost(\mathcal{C})=\#\{\text{negative edges inside clusters}\}+\#\{\text{positive edges between clusters}\},$$9

crossing the threshold

$$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$0

ATBOC is $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$1-Probably Correct and satisfies

$$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$2

which is within a factor of $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$3 of the lower bound. LUCBBOC retains the same stopping and declaration rules but replaces the expensive allocation computation by LUCB-style confidence bounds, yielding much lower computational cost and empirically comparable performance (Chandran et al., 20 Jan 2025).

For long-term AUV operations, OCF becomes a bounded-memory interpretation system over image streams. The seafloor-imagery framework combines a DPGMM backbone, online cluster merging, online cluster splitting, and representative-sample distillation. Clustering is triggered every $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$4 images, the latent dimension is $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$5, and representative memory is capped at $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$6 images. Merging uses both a Mahalanobis-distance criterion $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$7 and a covariance-volume criterion $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$8, with $$profit(\mathcal{C})=\#\{\text{positive edges inside clusters}\}+\#\{\text{negative edges between clusters}\},$$9 and $0.5$0 in the experiments; splitting is decided by recursive GMM fitting and AIC or BIC on representative samples. The framework attains the highest average F1 score of $0.5$1 across the three datasets among the comparative online approaches, with a standard deviation of $0.5$2 across three survey trajectories, while maintaining lower and bounded computational time as data volume increases. The best overall variant, OCF plus HKmeans sampling, reaches $0.5$3, whereas OCF plus density-based sampling provides the strongest accuracy-runtime compromise (Liang et al., 8 Sep 2025).

These task-specific systems show how OCF adapts to operational constraints that do not arise in abstract clustering theory: collar-based latency in diarization, sample complexity under fixed confidence in bandits, and constant-memory historical consolidation in long-duration marine robotics.

6. Boundaries, adjacent frameworks, and recurrent trade-offs

The literature also clarifies what OCF is not. OmniFC, despite being a sophisticated clustering framework, is explicitly “not an online clustering framework as defined in the paper.” It is a federated, model-agnostic, secure distance-reconstruction system in which clients share only encoded data, the server reconstructs the global distance matrix exactly when $0.5$4, and privacy is guaranteed against collusion of up to $0.5$5 clients (Yan et al., 19 May 2025). The distinction matters because “online,” “federated,” and “distributed” impose different algorithmic constraints.

A related but distinct streaming direction is online multi-view clustering with incomplete views. OMVC models large-scale incomplete multi-view clustering as a joint weighted nonnegative matrix factorization problem, processes data chunk by chunk to reduce memory requirement, learns latent feature matrices for all views and pushes them toward a consensus, introduces lasso regularization for robustness, uses dynamic weight setting to downweight missing instances, and accelerates optimization through projected gradient descent using Hessian matrices (Shao et al., 2016). This suggests an adjacent family of OCF-like systems in which the online constraint primarily concerns memory and incomplete modalities rather than irrevocable decisions or fixed-latency output.

Across all these variants, several trade-offs recur. In graph OCF, irrevocability creates hard competitive limits and strong asymmetry between maximization and minimization objectives (Mathieu et al., 2010, Chrobak et al., 2014). In probabilistic and prototype-based OCF, adaptivity in the number of clusters requires either posterior sampling, annealing schedules, or bifurcation control (Li et al., 2016, Mavridis et al., 2021). In deep and task-driven OCF, the dominant compromises are different: latency versus search depth in TBSC, sample complexity versus per-round optimization in ATBOC versus LUCBBOC, and clustering accuracy versus history replay cost in representative-sample seafloor clustering (Chen et al., 2022, Chandran et al., 20 Jan 2025, Liang et al., 8 Sep 2025).

A common misconception is therefore that OCF denotes a single canonical streaming clustering recipe. The literature instead supports a narrower and more technical characterization: OCF is a design principle for clustering under sequential revelation, bounded revision, and constrained computation. Its concrete form depends on what arrives online, what state may be retained, what actions are reversible, and what optimality notion—competitive ratio, regret, free energy, fixed-confidence sample complexity, DER, or F1—is treated as primary.