Iterative Latent Clustering (ILC)

• Iterative Latent Clustering (ILC) is a framework that iteratively assigns latent cluster labels using methods like expectation-maximization and seeded k-means to reveal underlying data structures.
• It refines latent feature embeddings and clustering assignments iteratively, providing both hard and soft groupings for applications in image segmentation, network analysis, and contingency table clustering.
• ILC methodologies ensure rapid convergence and interpretability by minimizing objective functions such as the k-means loss or KL divergence, supporting scalable and interactive analysis.

Iterative Latent Clustering (ILC) refers to a family of methods that iteratively assign latent cluster labels to observations by leveraging latent variable models and alternate optimization procedures. ILC methodologies are generally applied to high-dimensional data representations—either explicit, such as feature embeddings from neural networks, or implicit, such as entries in contingency tables or networks. The principal aim is to produce interpretable, data-driven soft or hard groupings reflecting the underlying data structure, using procedures grounded in expectation-maximization (EM), seed-augmented clustering, or similar iterative refinements (Chelebian et al., 2022, Bavaud, 2016).

1. Formal Definitions and Principal Variants

The ILC paradigm embodies a generic “embed, cluster, and refine” workflow in which the clustering takes place in a latent space. The two salient instantiations are:

• Seeded Iterative Clustering (SIC): Specialized for weakly-supervised region identification in large-scale image data, especially digital histopathology. Here, deep neural embeddings are clustered using constraints from sparse expert-provided seed points (Chelebian et al., 2022).
• Non-parametric Latent Modeling and Network Clustering: Focused on soft clustering (and co-clustering) of contingency tables and network data, via an alternating minimization of Kullback-Leibler (KL) divergence between empirical data and complete-data log-linear models, using EM-style updates (Bavaud, 2016).

While differing in statistical assumptions, both frameworks utilize alternating assignment-update cycles to minimize an application-specific objective, yielding either hard or soft assignments.

2. Seeded Iterative Clustering in Neural Latent Spaces

SIC targets semi-supervised, scalable segmentation or region annotation in whole-slide images. The procedure comprises the following key steps:

1. Latent Feature Extraction: Images are decomposed into overlapping patches (e.g., $256\times256$ px at 20$\times$ magnification), and embedded into vectors $x_i\in\mathbb{R}^d$ using CNN backbones (ResNet-18, ResNet-50; $d=512$–$2048$), pretrained or via self-supervised methods (SimCLR, HistoSSL). The patch embedding matrix $X$ is $N\times d$.
2. Clustering Objective: Binary hard assignments $r_{i,k}\in\{0,1\}$ minimize the k-means objective:

$$L_{\rm kmeans}(X,R,\mu) = \sum_{i=1}^N\sum_{k=1}^K r_{i,k} \|x_i - \mu_k\|_2^2$$

where $\mu_k$ denotes cluster centroids in latent space.

1. Seeded K-means Modifications:

Sparse user annotations determine $\times$0 seed sets. - Initial centroids are computed as means of seed embeddings: $\times$1. - Seed assignments are constrained: $\times$2 for $\times$3 and all $\times$4.

1. Iterative Refinement: The cycle “cluster $\times$5 restrict $\times$6 re-cluster” is repeated, with the clustering restricted to patches in the current positive cluster, until the F$\times$7-score on seed points ceases to improve.
2. Operational Properties:
   • Each clustering iteration is $\times$8; typical runs require $\times$9 iterations.
   • Patch-level F$x_i\in\mathbb{R}^d$0 for tumor/benign delineation, for example, ranges between $x_i\in\mathbb{R}^d$1–$x_i\in\mathbb{R}^d$2 depending on the embedding [see Table].

CNN Backbone	Embedding Type	Patch F₁ (Tumor/Benign)
ResNet-18	ImageNet	$x_i\in\mathbb{R}^d$3
ResNet-50	ImageNet	$x_i\in\mathbb{R}^d$4
ResNet-18	SimCLR	$x_i\in\mathbb{R}^d$5
ResNet-18	HistoSSL	$x_i\in\mathbb{R}^d$6

Convergence is typically achieved in 3–5 iterations and robust results are observed even with as few as 1–5 seeds per class, contingent on the discriminative capacity of the latent space (Chelebian et al., 2022).

3. EM-Based Iterative Latent Clustering for Contingency Tables and Networks

The EM-based ILC framework applies to tabular and network data via non-parametric latent variable models. The central elements are:

1. Data Representation and Model:
   • Data: $x_i\in\mathbb{R}^d$7 contingency table $x_i\in\mathbb{R}^d$8, $x_i\in\mathbb{R}^d$9, or an $d=512$0 adjacency matrix for networks.
   • Latent variable(s): discrete cluster indices $d=512$1 (or $d=512$2 for co-clustering).
   • Complete-data log-linear model:

   $d=512$3

2. KL Divergence Objective: Fit is measured by $d=512$4.

3. Alternating EM Updates:

   • E-step: Soft cluster responsibilities (posteriors) $d=512$5 assigned via

   $d=512$6

• M-step: Mixture and emission updates

  $d=512$7

Assignments remain soft throughout (i.e., fractional memberships).

1. Specializations:

   • Weighted Networks: Takes $d=512$8, $d=512$9, yielding symmetric cluster models for communities in graphs.
   • Co-clustering: Distinct latent variables for rows/columns; model extended for HMM-like bigram matrices.
2. Theoretical Properties: Each EM cycle reduces (or leaves unchanged) the KL objective, converging to a local minimum. All update steps admit closed forms, and convexity of the feasible spaces governs proper alternating projections (Csiszár–Tusnády theorem). Case studies include migration networks, text bigram clustering, and term-document matrices (Bavaud, 2016).

4. Comparative Analysis of Frameworks

Both SIC (Chelebian et al., 2022) and the EM-based latent models (Bavaud, 2016) instantiate the ILC paradigm by alternating latent reassignments and cluster updates within a probabilistic or geometric objective.

• Constraint Handling: SIC utilizes hard seed constraints for labeled points during k-means (semi-supervised), while EM-based models propagate responsibilities softmax-style.
• Data Domains: SIC is optimized for high-dimensional image patches; EM-based models address tabular, network, or sequence data.
• Cluster Assignment Type: Hard (SIC) versus soft (EM-ILC) assignments.
• Stopping Criteria: SIC halts upon annotation-fidelity (F$2048$0) non-improvement; EM-ILC relies on KL divergence decrease.

A plausible implication is that future ILC variants could hybridize hard and soft assignment mechanisms, or unify seed selection and model refinement in feedback-driven loops.

5. Scalability, Limitations, and Practical Considerations

SIC offers efficient post-embedding clustering, with critical computation decoupled from embedding generation (performed on GPU). Iterations are rapid (few seconds on CPU for $2048$1 matrices), and annotation effort is minimized due to interactive seeding.

The EM-based ILC is scalable to moderate-sized contingency tables and supports interpretability via soft memberships and marginal cluster assignment probabilities. Limiting factors include the local-minimum convergence of both procedures and, in the case of SIC, binary-class restriction (though extensions to multi-class seeding and soft/probabilistic k-means are noted).

In SIC, performance degrades when positive regions occupy a vanishingly small fraction (e.g., $2048$2) of total data (Chelebian et al., 2022). For EM-based ILC, categorical sequence inhomogeneity and assignment ambiguity are handled by model variants.

6. Research Directions and Extensions

Both ILC frameworks admit direct extensions:

• Multi-class extensions: SIC can generalize recursively or directly to $2048$3 via multiclass seeded k-means; EM-based ILC generalizes via increased latent dimension.
• Probabilistic or fuzzy assignment: SIC may incorporate seeded GMM or soft constraints; EM-ILC is already soft but could be adapted with additional priors.
• Alternative feedback and stopping: Objectives beyond F$2048$4 (intersection-over-union, balanced accuracy) or adaptive stopping can refine clustering outcomes.

The formalism affords application to transfer learning (by comparing neural embeddings across domains), large-scale weakly-supervised annotation, and interpretable network analysis.

7. Theoretical Guarantees and Empirical Observations

Both ILC instantiations guarantee non-increasing loss per iteration (F$2048$5 for SIC, KL for EM-ILC), with efficient closed-form iterative updates. SIC demonstrates competitive annotation accuracy with sparse seeds and rapid convergence, and EM-based ILC produces interpretable, domain-informed soft clusterings, evidenced in cases such as term-document and migration network analyses (Chelebian et al., 2022, Bavaud, 2016).

The ILC paradigm thus provides a unified blueprint for iterative, scalable assignment of latent structure, applicable across image, network, and tabular domains, permitting both user interaction and autonomous model improvement.