clustering analysis

Clustering high-dimensional data inherently involves a trade-off between abstraction—discarding redundant information—and representation—preserving discriminative structures. This work systematically examines the design principles and limitations of prevalent approaches, including K-means, subspace clustering, and deep clustering, and proposes a framework that explicitly decouples the latent space to separate clustering-relevant from clustering-irrelevant information. By doing so, the method enables a controllable balance between abstraction and representation. The proposed framework not only elucidates a key mechanism by which current deep clustering methods avoid degenerating into pure representation learning but also opens new avenues for developing more efficient and interpretable adaptive clustering algorithms.

This work addresses the challenge in unsupervised document clustering of simultaneously distinguishing document categories (e.g., invoices vs. purchase orders) and fine-grained templates within each category. We propose a multi-modal embedding-based fine-grained clustering framework that fuses representations from pre-trained models: textual (SBERT), layout-language (LayoutLMv1/v3), visual (DiT, Donut), and document-image alignment (ColPali). The resulting joint embeddings drive dual-level clustering—document-category and template-level—using k-Means and DBSCAN. To our knowledge, this is the first systematic empirical validation of multi-modal embeddings for unsupervised document organization, revealing the representational capabilities and limitations of diverse models for structured document understanding. Experiments demonstrate significant improvements in clustering accuracy and template discrimination over unimodal baselines. The framework provides a reusable pipeline and an empirical benchmark for intelligent document processing.

This paper addresses the challenge of evaluating and selecting appropriate clustering algorithms—K-means, DBSCAN, and spectral clustering—for high-dimensional data. We propose a unified unsupervised evaluation framework that integrates multiple dimensionality reduction techniques (PCA, t-SNE, UMAP) with composite validity metrics (silhouette coefficient, Adjusted Rand Index, Normalized Mutual Information). For the first time, we systematically uncover synergistic interactions between dimensionality reduction methods and clustering algorithms: UMAP preprocessing significantly enhances spectral clustering performance on complex manifold-structured data (e.g., MNIST, Fashion-MNIST, UCI HAR); K-means maintains superior computational efficiency; and DBSCAN excels at detecting irregularly shaped clusters. The framework provides reproducible, interpretable, and empirically grounded guidance for algorithm selection in high-dimensional clustering tasks.

Traditional correlation clustering struggles to model hierarchical structures and handle heterogeneous pairwise similarities—both positive and negative—within a unified framework. Method: We propose Hierarchical Correlation Clustering (HCC), the first correlation clustering formulation explicitly designed for hierarchical settings. HCC jointly optimizes cluster hierarchy and low-dimensional representations via an unsupervised tree-preserving embedding scheme. Furthermore, we generalize the minimax distance metric to correlation clustering, yielding a robust, noise-resilient dissimilarity measure tailored to heterogeneous pairwise constraints. Contribution/Results: Extensive experiments on standard benchmarks demonstrate that HCC significantly improves hierarchical clustering quality, embedding fidelity, and downstream task performance over state-of-the-art methods. By unifying hierarchy learning, representation learning, and robust constraint modeling in an unsupervised setting, HCC establishes a novel paradigm for unsupervised hierarchical representation learning.

Subspace clustering in high-dimensional data often yields multiple semantically distinct subspaces, yet existing methods require manual specification of both the number of subspaces and the number of clusters within each—rendering them parameter-sensitive and poorly interpretable. This paper proposes an automatic, non-redundant multi-subspace clustering framework. First, it introduces the Minimum Description Length (MDL) principle to non-redundant clustering, enabling joint, adaptive inference of both the optimal number of subspaces and the cluster count per subspace. Second, it designs a split-merge-based greedy search strategy coupled with a subspace-level outlier encoding mechanism, allowing simultaneous outlier detection. Evaluated on multiple benchmark datasets, the method achieves competitive accuracy against state-of-the-art approaches while significantly improving parameter robustness, model interpretability, and practical applicability.

Real-world data often exhibit hierarchical structure, yet existing deep hierarchical clustering methods suffer from poor scalability and limited performance. This paper proposes a lightweight, fine-tuning-free post-hoc framework that directly constructs high-quality hierarchical trees from logits produced by arbitrary pre-trained models—including unsupervised clusterers or ImageNet classifiers. Methodologically, it introduces two key components: (i) a logit-based spectral clustering variant, and (ii) a gradient-free hierarchical agglomerative algorithm that reconstructs similarity in feature space via logit distillation. Crucially, we provide the first theoretical and empirical evidence that logit distillation outperforms complex end-to-end hierarchical modeling. Our approach surpasses dedicated deep hierarchical clustering models across multiple benchmarks, reduces computational overhead by 10×, and—uniquely—enables通用, semantically consistent hierarchical discovery across both unsupervised and supervised settings.

Unsupervised feature selection (FS) faces a fundamental challenge: existing methods typically evaluate features in isolation, failing to capture semantically coherent feature groups—such as image pixel blocks or functionally connected brain regions—that exhibit synergistic discriminative power, and often rely on predefined groupings or label supervision. To address this, we propose GroupFS—the first end-to-end differentiable framework for unsupervised FS that jointly learns latent feature group structures and selects the most discriminative groups, without requiring prior grouping knowledge or labels. Its core innovation lies in integrating Laplacian smoothing constraints on both the feature graph and the sample graph, coupled with group-wise sparsity regularization, enabling gradient-based optimization of structured, compact representations. Evaluated on nine cross-domain benchmark datasets (image, tabular, and biological), GroupFS consistently improves clustering performance and identifies semantically meaningful, interpretable feature groups—advancing unsupervised FS toward structure-aware, end-to-end joint learning.

Existing clustering comparison methods struggle to effectively assess the agreement between clustering results containing overlapping clusters and outliers and ground-truth labels, often yielding misleading evaluations due to structural biases. This work presents the first systematic approach to measuring clustering similarity tailored for such complex scenarios, integrating set-matching principles with information-theoretic concepts. The proposed measure is rigorously defined and satisfies several desirable theoretical properties. Comprehensive theoretical analysis and extensive experiments demonstrate its superior robustness and fairness, significantly mitigating the evaluation bias inherent in conventional metrics when confronted with overlapping structures and outliers. This method thus provides a reliable tool for the quantitative comparison of complex clustering outcomes.

This study addresses the lack of systematic guidance in parameter selection and result evaluation for unsupervised data grouping methods by proposing SmartIterator, an exploratory framework grounded in a six-stage visual analytics pipeline. The approach uniquely treats the complete sequence of groupings generated through parameter sweeps as the primary analytical object, integrating quality metrics, stability assessments, member confidence scores, and domain context to deliver method-specific, actionable workflows for tasks such as clustering and topic modeling. Implemented via the IteraScope visualization system—which features semantic color encoding, group embeddings, Sankey transition flows, violin plots, and repeated prototype detection using HDBSCAN—the framework demonstrates its efficacy across three diverse datasets: social media, regional statistics, and academic publications, enabling analysts to comprehensively interpret data structures and make informed decisions.

Existing dimensionality reduction methods—such as t-SNE and UMAP—preserve local structure and yield well-separated clusters but often severely distort global geometry. To address this, we propose a modular visualization framework based on a “cluster-then-embed” paradigm: clustering (e.g., K-means, DBSCAN) and embedding (e.g., t-SNE, UMAP) are decoupled, and inter-cluster alignment constraints are explicitly imposed to govern the relative positions of local embeddings. This preserves local fidelity while restoring global structural consistency. The modular design enhances procedural transparency and result interpretability. Experiments on multiple synthetic and real-world datasets demonstrate that our method achieves visualization quality comparable to state-of-the-art approaches, while significantly improving global geometric fidelity and user interpretability.

Unsupervised image classification suffers from insufficient synergy between representation learning and clustering, with existing methods often neglecting feature diversity under frozen backbone models. This paper proposes a multi-head clustering framework featuring three key innovations: (1) adaptive nearest-neighbor selection to dynamically optimize similarity metrics; (2) a clustering ensemble mechanism that resolves inter-head conflicts to generate robust consensus pseudo-labels; and (3) a lightweight classifier training paradigm guided solely by pseudo-labels. Crucially, the method requires no fine-tuning of the backbone network. It achieves, for the first time in fully unsupervised settings, 70.4% top-1 accuracy on ImageNet, and 99.3% and 89.0% on CIFAR-10 and CIFAR-100, respectively. The approach outperforms prior work across all ten benchmark datasets, substantially narrowing the performance gap with supervised counterparts.