Amortized Clustering: Methods & Applications

• Amortized clustering is a family of methods that shifts computational effort offline through reusable inference maps or dynamic updates to efficiently partition data.
• It accommodates variable cluster numbers, maintains permutation invariance, and directly handles uncertainty in assignments across different data domains.
• Implementations such as amortized Bayesian inference and direct neural clustering demonstrate significant speedups, exemplified by reductions from 1969s to 184s for large-scale clustering tasks.

Searching arXiv for papers on amortized clustering and closely related formulations. Amortized clustering denotes a family of methods that shift computational effort away from solving a clustering problem from scratch for each dataset and toward reusable computation performed offline. In one dominant usage, the reused computation is a learned inference map: a neural network, prior-fitted network, or task-conditioned model is trained across many clustering problems and then applied to new data in a few forward passes or by direct posterior sampling. In a second, algorithmic usage, amortization refers to update complexity in dynamic clustering, where insertions and deletions are handled without recomputing a clustering from scratch. Across these literatures, the recurrent concerns are permutation invariance, variable numbers of clusters, uncertainty over assignments, and reuse of structure across related tasks or updates (Pakman et al., 2018, Lee et al., 2019, Wang et al., 2020, Bateni et al., 2021, Li et al., 2023, Zhao et al., 2024, Al-Jarrah et al., 16 Mar 2025, Xia et al., 1 Aug 2025, Zhao et al., 29 Jan 2026, López-Oriona et al., 13 May 2026).

1. Terminological scope and major formulations

The literature uses the term in several technically distinct but related senses. In amortized Bayesian clustering, the objective is to approximate posterior distributions over cluster assignments once and reuse that approximation across datasets. In direct neural clustering, the objective is to learn a dataset-to-partition map that replaces per-dataset optimization. In prior-fitted clustering, the model is pretrained on synthetic clustering tasks and deployed zero-shot. In dynamic clustering, amortization refers to update time rather than learned inference. Adjacent literatures also use clustering as the mechanism enabling amortization, for example by clustering a library of transport maps or clusters of locally optimal solutions (Pakman et al., 2018, Lee et al., 2019, Bateni et al., 2021, Li et al., 2023, Al-Jarrah et al., 16 Mar 2025, Zhao et al., 29 Jan 2026, López-Oriona et al., 13 May 2026).

Usage	Representative methods	Reused computation
Amortized posterior inference	NCP, CCP	learned conditional probabilities and posterior samples
Direct neural clustering	DAC	few forward passes on a new dataset
Prior-fitted clustering	TabClustPFN	synthetic-pretraining over clustering tasks
Feature-based amortized inference	time-series neural clustering	learned pairwise affinities from simulated data
Dynamic algorithmics	fully dynamic $k$-centers	amortized update time
Adjacent task-family reuse	A-OTF, AmorGS	clustered map libraries or clustered solution families

A common misconception is that amortized clustering denotes a single architecture. The published literature instead includes sequential posterior samplers for exchangeable mixture models, Set Transformer filtering networks, graph community detectors built from GNN backbones and amortized clustering heads, tabular prior-fitted networks that infer both assignments and cardinality, feature-based pairwise affinity learners for time series, and fully dynamic metric algorithms where the amortized object is update time rather than inference (Pakman et al., 2018, Lee et al., 2019, Wang et al., 2020, Bateni et al., 2021, Zhao et al., 29 Jan 2026, López-Oriona et al., 13 May 2026).

2. Amortized Bayesian inference over cluster assignments

A foundational probabilistic formulation is the Neural Clustering Process, which targets posterior inference in exchangeable clustering models such as Dirichlet process mixture models (Pakman et al., 2018). The model assumes

$$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$

and factorizes the clustering posterior sequentially as

$$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$

Amortization consists in learning the conditionals $p(c_n \mid c_{1:n-1}, \mathbf{x})$ once, so that future datasets can be processed without Gibbs sampling or variational optimization.

The technical core is a symmetry-invariant representation of partial clusterings. For each cluster,

$H_k = \sum_{i : c_i = k} h(x_i),$

for the set of occupied clusters,

$G = \sum_{k=1}^{K} g(H_k),$

and for the unassigned points,

$Q = \sum_{i=n+1}^{N} h(x_i).$

The learned conditional is then

$$p_{\theta}(c_n=k \mid c_{1:n-1}, \mathbf{x}) = \frac{\exp\{f(G_k,Q,h_n)\}} {\sum_{k'=1}^{K+1}\exp\{f(G_{k'},Q,h_n)\}},$$

with support for both assignment to an existing cluster and creation of a new cluster (Pakman et al., 2018). This construction is designed to respect permutations within clusters, permutations between clusters, and permutations of unassigned points.

The computational consequence is that a full sample of $c_{1:N}$ costs $O(NK)$, the same asymptotic cost as one Gibbs sweep, while the output samples are i.i.d. approximate posterior samples and can be generated in parallel on GPUs (Pakman et al., 2018). Training requires only samples from the generative model, so the same framework applies to both conjugate and non-conjugate settings. On a 1D comparison reported in the appendix, NCP took about 184 seconds for 20,000 samples, whereas Gibbs took about 1969 seconds for 20,000 samples plus 1000 burn-in iterations (Pakman et al., 2018).

This probabilistic line was extended to graphs in amortized community detection. There, a GNN encoder provides node embeddings $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$0, and an amortized clustering network such as NCP or CCP models the posterior over community assignments with variable $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$1 and parallel posterior sampling (Wang et al., 2020). The graph formulation emphasizes that amortized clustering is not restricted to Euclidean point clouds: the amortized object can be a posterior over partitions in a structured domain.

3. Sequential neural clustering and set-based architectures

Deep Amortized Clustering replaces per-dataset optimization with a learned, permutation-invariant clustering procedure that operates in a few forward passes (Lee et al., 2019). The method is meta-learned using labelled datasets, and the training distribution provides the prior knowledge that conventional algorithms usually encode manually through assumptions about cluster shapes or numbers.

DAC clusters one cluster at a time. The network repeatedly takes a set of points, predicts the parameters of one cluster, predicts which points belong to that cluster, removes those points, and repeats until all points are assigned (Lee et al., 2019). This filtering formulation is central to DAC’s ability to handle a variable number of clusters and to infer the cluster count automatically as the number of filtering iterations until completion.

The architecture is based on the Set Transformer. Its building blocks include the Multihead Attention Block,

$$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$2

the Self-Attention Block,

$$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$3

pooling by multihead attention,

$$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$4

and the Induced Self-Attention Block,

$$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$5

which reduces quadratic cost (Lee et al., 2019). The actual filtering network encodes the whole set, decodes one cluster representation $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$6, and then predicts a soft membership mask $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$7 conditioned on both the encoded points and that cluster representation.

Two label-invariant training objectives are central. Minimum Loss Filtering minimizes over the true clusters to avoid imposing an arbitrary label order; Anchored Filtering conditions on a random anchor point and asks the network to recover the anchor’s cluster (Lee et al., 2019). When an explicit density model is used, the objective combines a binary cross-entropy term for the mask with a likelihood term $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$8; when densities are hard to model, the likelihood term can be removed. The paper also describes density modeling with Masked Autoregressive Flow and VAE/Neural Statistician-style latent-variable models (Lee et al., 2019).

Empirically, DAC generalizes beyond the number of clusters seen during training. On 2D mixtures of Gaussians, it generalizes from training on up to 4 clusters to testing on up to 12 clusters; on Embedded ImageNet, it outperforms $$c_1,\ldots,c_N \sim p(c_{1:N}), \qquad \mu_k \sim p(\mu_k), \qquad x_i \sim p(x_i \mid \mu_{c_i}),$$9-means, spectral clustering, DEC, KCL, and MCL in the reported setup; on Omniglot it is competitive with strong transfer clustering baselines while being dramatically faster (Lee et al., 2019). This establishes a distinct sense of amortized clustering: a learned set-to-set clustering operator rather than an amortized posterior sampler.

4. Single-pass, prior-fitted, and domain-adapted systems

TabClustPFN extends amortized clustering to exploratory tabular data analysis by performing amortized Bayesian inference over both cluster assignments and cluster cardinality in a single forward pipeline (Zhao et al., 29 Jan 2026). Its posterior factorization is

$$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$0

and the architecture mirrors this decomposition with a Partition Inference Network approximating $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$1 and a Cardinality Inference Network approximating $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$2. The model is pretrained on synthetic datasets drawn from a flexible clustering prior. That prior mixes a GMM prior with probability 0.4 and a ZEUS prior with probability 0.6, with $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$3, $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$4, and $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$5 under a non-uniform prior favoring $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$6 (Zhao et al., 29 Jan 2026). PIN outputs soft memberships, CIN aggregates normalized Gram matrices

$$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$7

into permutation-invariant fingerprints, and the model clusters unseen datasets in a single forward pass without dataset-specific retraining or hyperparameter tuning (Zhao et al., 29 Jan 2026).

Domain-specialized variants illustrate how the same amortization principle is adapted to different data modalities. In graph community detection, a GNN backbone such as GraphSAGE or GatedGCN feeds an amortized clustering head such as NCP, CCP, or DAC, yielding posterior samples over community labels and uncertainty over the number of communities (Wang et al., 2020). The attention-enhanced variant GCN-CCP-Attn is reported as the best-performing model overall on SNAP datasets, and the amortized models are reported as more than 10× faster at inference than DMoN (Wang et al., 2020).

For time series, amortized clustering can be formulated as feature-based pairwise inference rather than direct assignment prediction. “Amortized Neural Clustering of Time Series based on Statistical Features” learns pairwise co-membership probabilities from autocorrelations and quantile autocorrelations using simulated datasets and binary cross-entropy on pairwise labels (López-Oriona et al., 13 May 2026). A shared embedding network processes each series, the two embeddings are combined by summation to ensure permutation invariance, and a post-processing graph method such as spectral clustering or Louvain recovers the partition (López-Oriona et al., 13 May 2026). In the GARCH(1,1) scenario, the reported mean/median ARI values are 0.490 / 0.431 for the proposed spectral method, 0.311 / 0.175 for the proposed Louvain method, 0.388 / 0.285 for $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$8-means, 0.312 / 0.202 for $$p(c_{1:N}\mid \mathbf{x}) = p(c_1\mid \mathbf{x})\, p(c_2\mid c_1,\mathbf{x}) \cdots p(c_N\mid c_{1:N-1},\mathbf{x}).$$9-medoids, and 0.409 / 0.307 for hierarchical clustering (López-Oriona et al., 13 May 2026).

Several adjacent systems use clustering as the mechanism that makes amortization feasible. A-OTF clusters pre-trained optimal transport maps with $p(c_n \mid c_{1:n-1}, \mathbf{x})$0-medoids and approximates the current map by a softmax-weighted linear combination of cluster medoid maps,

$p(c_n \mid c_{1:n-1}, \mathbf{x})$1

thereby avoiding online retraining at every time step (Al-Jarrah et al., 16 Mar 2025). AmorGS exploits clustering structure in locally optimal solutions, using a decoupled multimodal design with CVAE + GMM for time and mass variables and LSTM for the control variable (Li et al., 2023). In single-cell transcriptomics, an amortized stochastic variational Bayesian GPLVM uses an encoder

$p(c_n \mid c_{1:n-1}, \mathbf{x})$2

to produce a latent space that is cluster-friendly rather than to optimize cluster assignments explicitly (Zhao et al., 2024). In a Floquet-lattice application, amortized clustering denotes a task-conditioned algorithm selector over K-means++, hDBSCAN, agglomerative clustering, and mean-shift, trained across tasks indexed by the gain/loss parameter $p(c_n \mid c_{1:n-1}, \mathbf{x})$3 and evaluated by DBI, SC, and a time penalty (Xia et al., 1 Aug 2025).

5. Dynamic and streaming interpretations of amortization

A different but established usage of amortized clustering appears in fully dynamic clustering, where “amortized” refers to update time under insertions and deletions rather than to meta-learned inference (Bateni et al., 2021). The fully dynamic $p(c_n \mid c_{1:n-1}, \mathbf{x})$4-centers problem maintains a clustering of the current active set $p(c_n \mid c_{1:n-1}, \mathbf{x})$5 in an arbitrary metric space $p(c_n \mid c_{1:n-1}, \mathbf{x})$6, with objective

$p(c_n \mid c_{1:n-1}, \mathbf{x})$7

where $p(c_n \mid c_{1:n-1}, \mathbf{x})$8 is a set of at most $p(c_n \mid c_{1:n-1}, \mathbf{x})$9 centers (Bateni et al., 2021).

The main theorem of “Optimal Fully Dynamic $H_k = \sum_{i : c_i = k} h(x_i),$0-Centers Clustering” gives the first fully dynamic algorithm for $H_k = \sum_{i : c_i = k} h(x_i),$1-centers clustering in an arbitrary metric space that maintains an optimal $H_k = \sum_{i : c_i = k} h(x_i),$2 approximation in $H_k = \sum_{i : c_i = k} h(x_i),$3 amortized update time (Bateni et al., 2021). The paper states that previously the best known amortized update time was $H_k = \sum_{i : c_i = k} h(x_i),$4, and proves that any algorithm for $H_k = \sum_{i : c_i = k} h(x_i),$5-clustering tasks in arbitrary metric spaces, including $H_k = \sum_{i : c_i = k} h(x_i),$6-means, $H_k = \sum_{i : c_i = k} h(x_i),$7-medians, and $H_k = \sum_{i : c_i = k} h(x_i),$8-centers, must make at least $H_k = \sum_{i : c_i = k} h(x_i),$9 distance queries to achieve any non-trivial approximation factor (Bateni et al., 2021). In this sense, the paper closes the complexity of fully dynamic $G = \sum_{k=1}^{K} g(H_k),$0-centers clustering.

The technical route goes through dynamic maximal independent set maintenance: the core reduction is from dynamic $G = \sum_{k=1}^{K} g(H_k),$1-centers to maintaining a top-$G = \sum_{k=1}^{K} g(H_k),$2 lexicographically first maximal independent set in threshold graphs derived from the metric, via the Hochbaum–Shmoys reduction (Bateni et al., 2021). The same work gives a black-box transformation for metric spaces admitting locally sensitive hash functions. For Euclidean space, $G = \sum_{k=1}^{K} g(H_k),$3 spaces, the Hamming Metric, and the Jaccard Metric, for any $G = \sum_{k=1}^{K} g(H_k),$4, it yields a $G = \sum_{k=1}^{K} g(H_k),$5 approximate $G = \sum_{k=1}^{K} g(H_k),$6-centers solution in $G = \sum_{k=1}^{K} g(H_k),$7 amortized update time, simultaneously for all $G = \sum_{k=1}^{K} g(H_k),$8 (Bateni et al., 2021).

This algorithmic literature is conceptually related to neural amortization because both aim to avoid recomputation from scratch. The difference is the reused object. In learned amortization, one reuses parameters of an inference map. In dynamic clustering, one reuses a maintained combinatorial structure across updates (Bateni et al., 2021).

6. Recurring design principles, misconceptions, and limitations

Permutation invariance is a central design constraint throughout the field. NCP uses Deep Sets-style sums $G = \sum_{k=1}^{K} g(H_k),$9, $Q = \sum_{i=n+1}^{N} h(x_i).$0, and $Q = \sum_{i=n+1}^{N} h(x_i).$1 to remove dependence on point order and cluster labels (Pakman et al., 2018). DAC uses Set Transformer components together with minimum-over-clusters or anchored losses to avoid arbitrary label orderings (Lee et al., 2019). TabClustPFN uses prototype-based permutation-equivariant inference for assignments and sorted diagonal/off-diagonal statistics of Gram matrices for permutation-invariant cardinality inference (Zhao et al., 29 Jan 2026). Time-series amortized clustering uses summation of pair embeddings to guarantee symmetry of the pairwise co-membership predictor (López-Oriona et al., 13 May 2026). These constructions indicate that amortization by itself does not solve the clustering problem unless the network architecture respects the symmetries of partitions.

A second recurring principle is explicit handling of variable $Q = \sum_{i=n+1}^{N} h(x_i).$2. NCP lets each new point choose among existing clusters or create a new one (Pakman et al., 2018). DAC estimates the number of clusters by repeated filtering until no points remain (Lee et al., 2019). Graph community detection models are designed to avoid the fixed-softmax restriction of earlier GNN methods (Wang et al., 2020). TabClustPFN separates assignment inference from cluster-number inference through PIN and CIN (Zhao et al., 29 Jan 2026). The time-series framework delegates unknown-$Q = \sum_{i=n+1}^{N} h(x_i).$3 recovery to Louvain community detection (López-Oriona et al., 13 May 2026). A common misconception is therefore that amortized clustering merely accelerates fixed-$Q = \sum_{i=n+1}^{N} h(x_i).$4 algorithms; much of the literature instead treats unknown cardinality as a first-class inference target.

The principal limitations are equally recurrent. The methods often assume that training and test tasks come from the same or a related family. DAC is evaluated on new datasets coming from the same distribution used to generate training datasets (Lee et al., 2019). NCP is trained from samples of a specified generative model, so its quality depends on the fit of that model class (Pakman et al., 2018). TabClustPFN depends on the breadth of the synthetic prior, although its prior is deliberately flexible (Zhao et al., 29 Jan 2026). The time-series framework states that success requires the training simulator to reflect the real problem and that performance may degrade under severe misspecification (López-Oriona et al., 13 May 2026). A-OTF can only be as good as its offline library and clustering resolution, and the authors conjecture convergence with sufficiently dense training data but do not provide a proof (Al-Jarrah et al., 16 Mar 2025). AmorGS notes that data collection is expensive, that structure identification required human effort, and that generalizability is unclear (Li et al., 2023).

A further misconception is that amortized clustering always eliminates iterative computation. In practice, many systems still require structured downstream steps. DAC performs recursive cluster extraction (Lee et al., 2019). NCP samples labels sequentially point by point (Pakman et al., 2018). Graph community detection estimates a MAP sample from multiple posterior samples (Wang et al., 2020). TabClustPFN runs PIN over all candidate cluster counts before CIN selects $Q = \sum_{i=n+1}^{N} h(x_i).$5 (Zhao et al., 29 Jan 2026). Time-series methods still convert learned affinities into a partition through spectral clustering or Louvain (López-Oriona et al., 13 May 2026). Amortization therefore changes where computation occurs and what is learned, rather than removing algorithmic structure entirely.

Taken together, these literatures define amortized clustering as a general strategy for reusing structure across clustering problems. The reused structure may be a posterior factorization, a set-to-set neural procedure, a synthetic prior compiled into a prior-fitted network, a task-conditioned algorithm selector, a clustered library of expert maps, a family of solution manifolds, or a dynamic data structure maintained across updates (Pakman et al., 2018, Lee et al., 2019, Wang et al., 2020, Bateni et al., 2021, Li et al., 2023, Zhao et al., 2024, Al-Jarrah et al., 16 Mar 2025, Xia et al., 1 Aug 2025, Zhao et al., 29 Jan 2026, López-Oriona et al., 13 May 2026). The breadth of these formulations suggests that amortized clustering is best understood not as a single algorithm, but as a unifying computational principle for clustering under repeated, related inference tasks.