C$^{2}$TC: A Training-Free Framework for Efficient Tabular Data Condensation

Sijia Xu; Fan Li; Xiaoyang Wang; Zhengyi Yang; Xuemin Lin

cs.LGFeb 25, 2026

C $^{2}$ TC: A Training-Free Framework for Efficient Tabular Data Condensation

Authors: Sijia Xu, Fan Li, Xiaoyang Wang, Zhengyi Yang, Xuemin Lin

Organizations: University of New South Wales, Australia · Shanghai Jiao Tong University, China

Abstract

Tabular data is the primary data format in industrial relational databases, underpinning modern data analytics and decision-making. However, the increasing scale of tabular data poses significant computational and storage challenges to learning-based analytical systems. This highlights the need for data-efficient learning, which enables effective model training and generalization using substantially fewer samples. Dataset condensation (DC) has emerged as a promising data-centric paradigm that synthesizes small yet informative datasets to preserve data utility while reducing storage and training costs. However, existing DC methods are computationally intensive due to reliance on complex gradient-based optimization. Moreover, they often overlook key characteristics of tabular data, such as heterogeneous features and class imbalance. To address these limitations, we introduce C $^{2}$ TC (Class-Adaptive Clustering for Tabular Condensation), the first training-free tabular dataset condensation framework that jointly optimizes class allocation and feature representation, enabling efficient and scalable condensation. Specifically, we reformulate the dataset condensation objective into a novel class-adaptive cluster allocation problem (CCAP), which eliminates costly training and integrates adaptive label allocation to handle class imbalance. To solve the NP-hard CCAP, we develop HFILS, a heuristic local search that alternates between soft allocation and class-wise clustering to efficiently obtain high-quality solutions. Moreover, a hybrid categorical feature encoding (HCFE) is proposed for semantics-preserving clustering of heterogeneous discrete attributes. Extensive experiments on 10 real-world datasets demonstrate that C $^{2}$ TC improves efficiency by at least 2 orders of magnitude over state-of-the-art baselines, while achieving superior downstream performance.

Figures & tables

Panel of Figure 1 — Fig. 1 : Comparison of label allocation strategies.

Fig. 2 : Overall framework of

\mathrm{C^{2}TC}

Dataset	Size	#Num	#Cat	MinMaxCF	#Classes
Electricity (EL)	38,474	6	1	1.00	2
Adult (AD)	48,842	6	8	0.31	2
Jannis (JA)	83,733	54	0	0.04	4
Diabetes130US (DA)	101,766	13	36	0.21	3
Road Safety (RS)	111,762	29	3	1.00	2
Epsilon (EP)	500,000	2,000	0	1.00	2

TABLE I : The statistics of datasets. Num and Cat stands for numerical and categorical features, respectively. MinMaxCF denotes the ratio of min–max class frequency [ 64 ] .

Dataset	r	Random		Herding		K-Center		DM		GM		MTT		$\text{C}^{2}\text{TC}$		Whole Dataset
Dataset	r	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1	Acc	Macro-F1
EL	1%	57.0 $\pm$ 2.5	56.0 $\pm$ 3.4	54.1 $\pm$ 0.1	54.1 $\pm$ 0.1	53.8 $\pm$ 1.5	48.5 $\pm$ 5.6	57.4 $\pm$ 2.0	56.4 $\pm$ 3.0	50.6 $\pm$ 1.3	37.0 $\pm$ 7.3	58.0 $\pm$ 2.2	57.1 $\pm$ 2.9	66.1 $\pm$ 0.5	65.9 $\pm$ 0.5	76.5 $\pm$ 0.9	76.4 $\pm$ 0.9
	0.1%	53.5 $\pm$ 2.3	53.0 $\pm$ 2.5	54.2 $\pm$ 0.1	54.2 $\pm$ 0.1	50.6 $\pm$ 3.7	48.4 $\pm$ 4.3	51.6 $\pm$ 2.4	51.0 $\pm$ 2.4	50.3 $\pm$ 2.7	44.3 $\pm$ 7.2	52.6 $\pm$ 3.4	51.9 $\pm$ 3.4	64.5 $\pm$ 0.5	63.5 $\pm$ 0.6
	0.01%	51.2 $\pm$ 1.7	48.7 $\pm$ 4.5	53.7 $\pm$ 0.9	51.8 $\pm$ 4.7	53.7 $\pm$ 0.9	51.8 $\pm$ 4.7	50.9 $\pm$ 2.3	47.2 $\pm$ 3.2	50.3 $\pm$ 1.0	34.5 $\pm$ 4.3	51.0 $\pm$ 2.1	47.0 $\pm$ 3.7	62.9 $\pm$ 0.2	60.4 $\pm$ 2.7
AD	1%	77.2 $\pm$ 1.1	56.3 $\pm$ 6.1	72.9 $\pm$ 1.8	65.3 $\pm$ 1.1	76.1 $\pm$ 0.0	43.2 $\pm$ 0.0	78.5 $\pm$ 0.6	61.3 $\pm$ 3.4	76.1 $\pm$ 0.0	43.2 $\pm$ 0.0	78.5 $\pm$ 0.5	61.3 $\pm$ 2.9	81.7 $\pm$ 0.4	74.9 $\pm$ 0.7	85.5 $\pm$ 0.1	78.4 $\pm$ 0.4
	0.1%	75.7 $\pm$ 1.3	56.3 $\pm$ 5.8	67.8 $\pm$ 2.7	61.4 $\pm$ 1.0	76.0 $\pm$ 0.1	43.2 $\pm$ 0.0	76.3 $\pm$ 1.7	55.4 $\pm$ 7.6	76.1 $\pm$ 0.0	43.2 $\pm$ 0.0	76.2 $\pm$ 1.8	56.0 $\pm$ 7.6	81.5 $\pm$ 0.6	70.8 $\pm$ 1.1

TABLE II : Overall Performance Comparison. Whole Dataset denotes training on the full dataset and serves as an approximate upper bound. Bold indicates the best performance and underline means the runner-up.

Dataset	Method	MLP-Based			Transformer-Based			Avg.
Dataset	Method	MLP	MLP_PLR	RealMLP	FT_T	TabNet	TabR	Avg.
EL	Random	53.5	55.8	52.3	53.9	50.3	52.2	53.0
	DM	51.6	55.1	55.1	57.2	49.3	53.0	53.6
	GM	50.3	50.6	50.3	50.6	50.3	50.6	50.5
	MTT	52.6	57.0	52.7	56.0	51.1	53.9	53.9
	$\text{C}^{2}\text{TC}$	64.5	67.5	62.8	64.5	51.0	61.3	62.0

TABLE III : Generalization of DC methods across deep tabular learning models. Avg. stands for the average test accuracy of MLP, MLP_PLR, RealMLP, FT_T, TabNet, and TabR. Results for DM, GM, and MTT on EP, and GM on MI, are omitted as they fail to scale to these datasets due to time or memory constraints. All results are reported in terms of accuracy.

Fig. 3 : Condensation time comparison.

Panel of Figure 7 — Fig. 4 : Condensation time by varying $N$ and $F$ .

Panel of Figure 8 — Fig. 5 : Training time on condensed vs. whole datasets.

Strategy	EL		AD		EP		CO		MI
Strategy	Acc	MF1	Acc	MF1	Acc	MF1	Acc	MF1	Acc	MF1
Ratio	63.0	62.9	80.4	69.7	69.9	69.0	65.6	33.1	53.5	20.9
FIPC	63.0	62.9	74.2	70.7	69.9	69.0	47.3	40.5	34.2	20.8
$\text{C}^{2}\text{TC}$	64.5	63.5	81.5	70.8	75.2	74.8	69.6	46.4	53.6	21.5

TABLE IV : Comparison of label allocation strategies by Accuracy (Acc) and Macro-F1 (MF1).

Modules	Metrics	AD	JA	DA	AI	CO	MI
w/o Soft Step Size	Acc	74.5	50.9	36.8	–	60.9	46.6
w/o Soft Allocation	Acc	–	51.7	50.5	–	58.3	49.8
$\text{C}^{2}\text{TC}$	Acc	81.5	52.8	51.8	61.5	69.6	53.6
w/o Scale Factor	Macro-F1	69.0	29.2	32.6	–	35.8	17.0
$\text{C}^{2}\text{TC}$	Macro-F1	70.8	35.3	34.4	58.4	46.4	21.5

TABLE V : Ablation study of

\text{C}^{2}\text{TC}

modules.

Panel of Figure 11 — Fig. 6 : Parameter sensitivity analysis of $\gamma$ and $l$ .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	C 2 TC vs. Herding		C 2 TC vs. MTT
Dataset	Acc $p$ -val	MF1 $p$ -val	Acc $p$ -val	MF1 $p$ -val
EL	6.90 $\times$ 10 -33	7.34 $\times$ 10 -30	3.78 $\times$ 10 -15	8.79 $\times$ 10 -15
AD	4.01 $\times$ 10 -17	3.20 $\times$ 10 -20	5.37 $\times$ 10 -13	5.13 $\times$ 10 -10
JA	4.76 $\times$ 10 -1	1.44 $\times$ 10 -6	3.75 $\times$ 10 -3	2.48 $\times$ 10 -2
DA	4.45 $\times$ 10 -9	1.05 $\times$ 10 -13	9.07 $\times$ 10 -8	1.78 $\times$ 10 -12
RS	2.28 $\times$ 10 -2	5.56 $\times$ 10 -2	2.18 $\times$ 10 -2	3.60 $\times$ 10 -2

Appendix

TABLE VI : Paired t-test results (

p

-values) comparing

\text{C}^{2}\text{TC}

against competitive baselines (Herding and MTT). Bold values indicate statistical significance (

p<0.05

). Results for MTT on EP are omitted due to OOT.

Dataset	Label		One-hot + PCA		Target		HCFE
Dataset	Acc	MF1	Acc	MF1	Acc	MF1	Acc	MF1
AD	67.7	62.9	74.4	70.0	78.8	69.9	81.5	70.8
DA	50.2	24.3	51.3	30.8	47.1	33.5	51.8	34.4
AI	53.1	52.0	61.0	56.0	59.1	58.2	61.5	58.4

Appendix

TABLE VII : Comparison of categorical feature encoding strategies for data condensation.

Panel of Figure 14 — Fig. 7 : Impact of categorical feature imbalance.

Dataset Condensation

Dataset condensation creates tiny training sets that preserve the learning value of much larger datasets.

See all 19 papers

Sep 12, 2025

A Discrepancy-Based Perspective on Dataset Condensation

Given a dataset of finitely many elements

\mathcal{T} = \{\mathbf{x}_i\}_{i = 1}^N

, the goal of dataset condensation (DC) is to construct a synthetic dataset

\mathcal{S} = \{\tilde{\mathbf{x}}_j\}_{j = 1}^M

which is significantly smaller (

M \ll N

) such that a model trained from scratch on

\mathcal{S}

achieves comparable or even superior generalization performance to a model trained on

\mathcal{T}

. Recent advances in DC reveal a close connection to the problem of approximating the data distribution represented by

\mathcal{T}

with a reduced set of points. In this work, we present a unified framework that encompasses existing DC methods and extend the task-specific notion of DC to a more general and formal definition using notions of discrepancy, which quantify the distance between probability distribution in different regimes. Our framework broadens the objective of DC beyond generalization, accommodating additional objectives such as robustness, privacy, and other desirable properties.

Tong Chen, Raghavendra Selvan

Department of Computer Science, University of Copenhagen, Denmark

Distribution Matching Dataset Distillation Dataset Condensation

Ask PDF

Apr 13, 2026

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

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.

Mingjie Zhao, Yunfan Zhang, Yiqun Zhang +1

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

Clustering Deep Clustering Tabular Representation Learning Contrastive Learning

Ask PDF

Mar 16, 2026

TabKD: Tabular Knowledge Distillation through Interaction Diversity of Learned Feature Bins

Data-free knowledge distillation enables model compression without original training data, critical for privacy-sensitive tabular domains. However, existing methods does not perform well on tabular data because they do not explicitly address feature interactions, the fundamental way tabular models encode predictive knowledge. We identify interaction diversity, systematic coverage of feature combinations, as an essential requirement for effective tabular distillation. To operationalize this insight, we propose TabKD, which learns adaptive feature bins aligned with teacher decision boundaries, then generates synthetic queries that maximize pairwise interaction coverage. Across 4 benchmark datasets and 4 teacher architectures, TabKD achieves highest student-teacher agreement in 14 out of 16 configurations, outperforming 5 state-of-the-art baselines. We further show that interaction coverage strongly correlates with distillation quality, validating our core hypothesis. Our work establishes interaction-focused exploration as a principled framework for tabular model extraction.

Shovon Niverd Pereira, Krishna Khadka, Yu Lei

Department of Computer Science and Engineering, The University of Texas at Arlington

Feature Interaction Modeling Data-Free Knowledge Distillation Tabular ML Teacher-Student Learning Knowledge Distillation

Ask PDF

Abstract

Dataset Condensation

Explore similar work

A Discrepancy-Based Perspective on Dataset Condensation

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

TabKD: Tabular Knowledge Distillation through Interaction Diversity of Learned Feature Bins