Beyond Statistical Co-occurrence: Unlocking Intrinsic Semantics for Tabular Data Clustering

Mingjie Zhao; Yunfan Zhang; Yiqun Zhang; Yiu-ming Cheung

cs.AIApr 13, 2026

Beyond Statistical Co-occurrence: Unlocking Intrinsic Semantics for Tabular Data Clustering

Authors: Mingjie Zhao, Yunfan Zhang, Yiqun Zhang, Yiu-ming Cheung

Organizations: Department of Computer Science, Hong Kong Baptist University, Hong Kong, China · Department of Computer Science, Guangdong University of Technology, Guangzhou, China

Abstract

Deep Clustering (DC) has emerged as a powerful tool for tabular data analysis in real-world domains like finance and healthcare. However, most existing methods rely on data-level statistical co-occurrence to infer the latent metric space, often overlooking the intrinsic semantic knowledge encapsulated in feature names and values. As a result, semantically related concepts like Flu' and Cold' are often treated as symbolic tokens, causing conceptually related samples to be isolated. To bridge the gap between dataset-specific statistics and intrinsic semantic knowledge, this paper proposes Tabular-Augmented Contrastive Clustering (TagCC), a novel framework that anchors statistical tabular representations to open-world textual concepts. Specifically, TagCC utilizes Large Language Models (LLMs) to distill underlying data semantics into textual anchors via semantic-aware transformation. Through Contrastive Learning (CL), the framework enriches the statistical tabular representations with the open-world semantics encapsulated in these anchors. This CL framework is jointly optimized with a clustering objective, ensuring that the learned representations are both semantically coherent and clustering-friendly. Extensive experiments on benchmark datasets demonstrate that TagCC significantly outperforms its counterparts.

Figures & tables

Figure 1: Comparison between statistical encoding (Left) and semantic relationship (Right). Statistical encoding strategies treat related concepts like ‘Flu’ and ‘Cold’ as orthogonal to ‘Fracture’, i.e., equidistant. In contrast, ideally, the latent space should be aligned with open-world semantics to ensure that conceptually similar samples are partitioned into coherent clusters.

Figure 2: Overview of the TagCC framework. To bridge the semantic gap in tabular clustering, the framework augments raw table rows with LLM-synthesized semantic anchors $t_{i}$ . These anchors and raw features $\mathbf{x}_{i}$ are then processed by a dual-branch encoder, where the trainable tabular network is optimized to align with the frozen semantic backbone. The optimization is conducted on a unit hypersphere $\mathcal{Z}$ , where the model jointly minimizes CL loss $\mathcal{L}_{align}$ for semantic consistency and DC loss $\mathcal{L}_{proto}$ for structural compactness.

No.	Dataset	Abbrev.	Domain	$m$	$n$	$k^{*}$
1	Soybean	SO	Agriculture	35	307	15
2	Breast cancer	BR	Healthcare	9	286	2
3	Car evaluation	CA	Transportation	6	1728	4
4	Lenses	LE	Healthcare	4	24	3
5	Mammographic	MA	Healthcare	5	961	2
6	Zoo	ZO	Biology	15	101	7

Table 1: Dataset statistics.

m

n

, and

k^{*}

are the numbers of features, samples, and true number of clusters, respectively.

Data	Index	Kmeans	DEC	CC	SCARF	QGRL	IDC	TabDC	TagCC	$\Delta$
Data	Index	Kmeans	[ICLR’16]	[AAAI’21]	[ICLR’22]	[SIGKDD’24]	[ICML’24]	[SIGMOD’25]	[Ours]	$\Delta$
SO	ARI	0.3918 $\pm$ 0.001	0.3808 $\pm$ 0.032	0.3050 $\pm$ 0.012	0.4015 $\pm$ 0.013	0.2620 $\pm$ 0.042	0.1489 $\pm$ 0.031	0.3340 $\pm$ 0.019	0.4773 $\pm$ 0.029	$\uparrow$ 18.88%
	ACC	0.5694 $\pm$ 0.001	0.5621 $\pm$ 0.039	0.4637 $\pm$ 0.020	0.5324 $\pm$ 0.019	0.3906 $\pm$ 0.031	0.2907 $\pm$ 0.045	0.4990 $\pm$ 0.018	0.6101 $\pm$ 0.034	$\uparrow$ 7.15%
	NMI	0.6689 $\pm$ 0.001	0.6679 $\pm$ 0.024	0.5842 $\pm$ 0.013	0.6759 $\pm$ 0.009	0.4265 $\pm$ 0.049	0.3533 $\pm$ 0.047	0.6440 $\pm$ 0.012	0.7122 $\pm$ 0.015	$\uparrow$ 5.37%
BR	ARI	-0.0035 $\pm$ 0.000	0.0438 $\pm$ 0.058	0.0364 $\pm$ 0.033	-0.0033 $\pm$ 0.000	0.1700 $\pm$ 0.000	0.0564 $\pm$ 0.051	0.0150 $\pm$ 0.037	0.1489 $\pm$ 0.053	$\downarrow$ 12.41%
	ACC	0.5126 $\pm$ 0.000	0.5818 $\pm$ 0.083	0.5852 $\pm$ 0.051	0.5091 $\pm$ 0.003	0.7203 $\pm$ 0.000	0.6077 $\pm$ 0.063	0.5460 $\pm$ 0.068	0.7238 $\pm$ 0.015	$\uparrow$ 0.49%

Table 2: Clustering performance evaluated by the ACC, NMI, and ARI on various datasets (mean

\pm

std). The best and second-best results on each dataset are highlighted in bold and underlined , respectively.

\Delta

denotes the relative performance gain of the best method over the second-best, calculated as

(1st-2nd)/2nd\times 100\%

Data	ARI				ACC				NMI
Data	TTC	TLC	TCC	TagCC	TTC	TLC	TCC	TagCC	TTC	TLC	TCC	TagCC
SO	0.41 $\pm$ 0.04	0.10 $\pm$ 0.02	0.38 $\pm$ 0.03	0.48 $\pm$ 0.03	0.55 $\pm$ 0.04	0.28 $\pm$ 0.03	0.51 $\pm$ 0.04	0.61 $\pm$ 0.03	0.66 $\pm$ 0.03	0.20 $\pm$ 0.04	0.53 $\pm$ 0.04	0.71 $\pm$ 0.01
BR	0.02 $\pm$ 0.05	0.02 $\pm$ 0.03	0.07 $\pm$ 0.03	0.15 $\pm$ 0.05	0.56 $\pm$ 0.06	0.59 $\pm$ 0.06	0.67 $\pm$ 0.03	0.72 $\pm$ 0.02	0.02 $\pm$ 0.03	0.01 $\pm$ 0.01	0.03 $\pm$ 0.02	0.07 $\pm$ 0.03
CA	0.05 $\pm$ 0.03	0.02 $\pm$ 0.02	0.06 $\pm$ 0.05	0.19 $\pm$ 0.03	0.37 $\pm$ 0.04	0.45 $\pm$ 0.09	0.46 $\pm$ 0.12	0.52 $\pm$ 0.07	0.07 $\pm$ 0.04	0.03 $\pm$ 0.02	0.04 $\pm$ 0.02	0.24 $\pm$ 0.07
LE	0.08 $\pm$ 0.12	0.12 $\pm$ 0.13	0.21 $\pm$ 0.18	0.41 $\pm$ 0.15	0.50 $\pm$ 0.09	0.57 $\pm$ 0.07	0.62 $\pm$ 0.11	0.70 $\pm$ 0.10	0.22 $\pm$ 0.18	0.16 $\pm$ 0.10	0.27 $\pm$ 0.13	0.42 $\pm$ 0.02
MA	0.38 $\pm$ 0.02	0.24 $\pm$ 0.09	0.40 $\pm$ 0.01	0.41 $\pm$ 0.02	0.81 $\pm$ 0.01	0.74 $\pm$ 0.05	0.80 $\pm$ 0.00	0.82 $\pm$ 0.01	0.32 $\pm$ 0.02	0.19 $\pm$ 0.07	0.31 $\pm$ 0.01	0.34 $\pm$ 0.01

Table 3: Ablation studies of TagCC. TCC removes the semantic-aware transformation. TLC and TTC denote single-branch deep clustering using only textual or tabular embeddings, respectively. The best results are highlight in bold .

\overline{\textit{AR}}

reports Average Rank of each method.

Figure 3: t-SNE visualization of representations learned by SCARF, TabDC, and TagCC on the MU dataset. Orange and blue points mark the samples with the “true” cluster labels.

Figure 4: Convergence analysis on CA (left) and MA (right). The vertical dashed line at epoch 50 separates the semantic warm-up stage from the structure refinement stage.

Clustering

Clustering groups data points by shared structure, developing methods to discover, validate, and interpret groups across diverse data and constraints.

See all 345 papers

GRACE: LLM-Grounded Semantic Metric Spaces for Scalable Mixed-Data Clustering
Aug 8, 2026Zihua Yang, Zhencheng Xie, Junyang Chen +4Clustering Unsupervised Clustering
Guangdong University of Technology · Tsinghua University · Peking University +2
Enhancing Tabular Learners with Context-Aware Semantic Embeddings
Aug 4, 2026Günther Schindler, Maximilian Schambach, Johannes HöhneTabular Foundation Models Representation Learning
SAP SE
TabClustPFN: A Prior-data Fitted Network for Tabular Data Clustering
Jan 29, 2026Tianqi Zhao, Guanyang Wang, Yan Shuo Tan +1Cluster Number Estimation Clustering
Renmin University of China · Rutgers University · National University of Singapore +1

Aug 8, 2026J/K move · Enter open · S save

GRACE: LLM-Grounded Semantic Metric Spaces for Scalable Mixed-Data Clustering

Zihua Yang, Zhencheng Xie, Junyang Chen, Liang Xie+3

Guangdong University of Technology · Tsinghua University · Peking University+2

Clustering Unsupervised Clustering Metric Learning Tabular ML

Open paper AskPDF arXiv

Clustering mixed tabular data requires a unified metric space to bridge the inherent heterogeneity between continuous numerical measurements and discrete categorical symbols. Traditionally, algorithms rely entirely on dataset-internal statistics to estimate categorical relationships, which confines the learned metric to empirical co-occurrences and ignores conceptually obvious yet statistically unobserved affinities. Although LLMs offer external world knowledge, applying their text-centric reasoning to highly abstract tabular concepts presents significant challenges. Bridging this modality gap to construct a semantically complete metric typically requires embedding LLMs into iterative metric learning loops to dynamically optimize cross-modality representations. This incurs intractable computational overhead, forcing a compromise between semantic enrichment and scalability. Therefore, we propose GRACE, an LLM-grounded framework for scalable mixed-data clustering. GRACE shifts semantic acquisition to the attribute-value level via a multi-perspective LLM querying strategy, mapping heterogeneous values into knowledge-informed descriptions. Crucially, this one-shot grounding extracts general-purpose semantic representations that embed heterogeneous attributes into a unified space, decoupling expensive LLM invocation from iterative optimization. Furthermore, GRACE cross-validates these external semantics against dataset-internal statistical evidence to ensure alignment with the dataset-specific cluster structure. Ultimately, GRACE matches the scalability of conventional statistics-driven baselines while achieving superior clustering accuracy and conceptual interpretability over 11 competing methods. The source code is available at https://github.com/develop-yang/GRACE-GRACE-A

Abstract

Clustering

Explore similar work

GRACE: LLM-Grounded Semantic Metric Spaces for Scalable Mixed-Data Clustering

Enhancing Tabular Learners with Context-Aware Semantic Embeddings

TabClustPFN: A Prior-data Fitted Network for Tabular Data Clustering