XGenBoost: Synthesizing Small and Large Tabular Datasets with XGBoost

Jim Achterberg; Marcel Haas; Bram van Dijk; Marco Spruit

cs.LGMar 6, 2026

XGenBoost: Synthesizing Small and Large Tabular Datasets with XGBoost

Authors: Jim Achterberg, Marcel Haas, Bram van Dijk, Marco Spruit

Organizations: Public Health & Primary Care, Leiden University Medical Center, Leiden, The Netherlands · Research and Development, Statistics Netherlands (CBS), The Hague, The Netherlands · Leiden Institute of Advanced Computer Science, Leiden University, Leiden, The Netherlands

Abstract

Tree ensembles such as XGBoost are often preferred for discriminative tasks in mixed-type tabular data, due to their inductive biases, minimal hyperparameter tuning, and training efficiency. We argue that these qualities, when leveraged correctly, can make for better generative models as well. As such, we present XGenBoost, a pair of generative models based on XGBoost: i) a Denoising Diffusion Implicit Model (DDIM) with XGBoost as score-estimator suited for smaller datasets, and ii) a hierarchical autoregressive model whose conditionals are learned via XGBoost classifiers, suited for large-scale tabular synthesis. The architectures follow from the natural constraints imposed by tree-based learners, e.g., in the diffusion model, combining Gaussian and multinomial diffusion to leverage native categorical splits and avoid one-hot encoding while accurately modelling mixed data types. In the autoregressive model, we use a fixed-order factorization, a hierarchical classifier to impose ordinal inductive biases when modelling numerical features, and de-quantization based on empirical quantile functions to model the non-continuous nature of most real-world tabular datasets. Through two benchmarks, one containing smaller and the other larger datasets, we show that our proposed architectures outperform previous neural- and tree-based generative models for mixed-type tabular synthesis at lower training cost.

Figures & tables

metric	Shape	Trend	Detection	$\alpha$ -Precision	$\beta$ -Recall	MLE	DCR	Training Time	Inference Time
SMOTE	$6.654_{\pm 2.756}$	$5.154_{\pm 3.233}$	$6.885_{\pm 2.805}$	$7.769_{\pm 3.491}$	$3.462_{\pm 3.289}$	$\underline{4.808}_{\pm 3.805}$	$10.923_{\pm 2.637}$	-	$2.769_{\pm 1.177}$
ARF	$7.852_{\pm 2.641}$	$9.704_{\pm 1.938}$	$9.222_{\pm 2.044}$	$5.333_{\pm 3.063}$	$10.630_{\pm 0.742}$	$10.037_{\pm 2.457}$	$3.278_{\pm 1.734}$	$\underline{2.074}_{\pm 1.072}$	$3.926_{\pm 0.829}$
UT	$7.222_{\pm 2.532}$	$7.148_{\pm 2.627}$	$7.037_{\pm 2.993}$	$6.741_{\pm 3.046}$	$7.407_{\pm 1.474}$	$7.370_{\pm 2.452}$	$8.093_{\pm 3.000}$	$5.407_{\pm 0.888}$	$12.481_{\pm 1.221}$
FD	$9.407_{\pm 1.623}$	$5.407_{\pm 2.859}$	$8.407_{\pm 3.165}$	$7.000_{\pm 3.563}$	$8.481_{\pm 2.190}$	$5.333_{\pm 3.026}$	$6.648_{\pm 2.269}$	$10.630_{\pm 1.573}$	$7.185_{\pm 0.483}$
FF	$7.444_{\pm 2.326}$	$5.000_{\pm 2.617}$	$6.556_{\pm 3.055}$	$6.926_{\pm 2.401}$	$\underline{3.185}_{\pm 2.573}$	$5.037_{\pm 2.579}$	$9.278_{\pm 2.419}$	$10.074_{\pm 0.958}$	$5.593_{\pm 0.636}$
CTGAN	$12.407_{\pm 0.694}$	$12.296_{\pm 0.609}$	$12.296_{\pm 0.775}$	$11.815_{\pm 1.962}$	$12.630_{\pm 0.451}$	$12.667_{\pm 0.555}$	$\underline{2.685}_{\pm 2.284}$	$3.741_{\pm 0.594}$	$\mathbf{1.630}_{\pm 0.629}$

Table 1: Small Benchmark: metric ranks

(\mu\pm\sigma)

across the 27 datasets of the Small Benchmark. Best results per metric are in bold , second best underlined .

metric	Shape	Trend	Detection	$\alpha$ -Precision	$\beta$ -Recall	MLE	DCR	Training Time	Inference Time
SMOTE	$3.889_{\pm 0.782}$	$3.111_{\pm 0.782}$	$3.444_{\pm 0.726}$	$5.667_{\pm 1.323}$	$\mathbf{1.778}_{\pm 1.563}$	$\mathbf{2.111}_{\pm 1.364}$	$6.778_{\pm 0.441}$	-	$4.556_{\pm 0.882}$
ARF	$3.818_{\pm 0.982}$	$6.364_{\pm 1.206}$	$4.727_{\pm 0.467}$	$2.636_{\pm 1.027}$	$4.545_{\pm 1.508}$	$2.909_{\pm 1.514}$	$5.727_{\pm 0.786}$	$3.818_{\pm 0.874}$	$6.000_{\pm 0.775}$
CTGAN	$6.000_{\pm 1.000}$	$5.273_{\pm 0.905}$	$6.455_{\pm 0.688}$	$5.636_{\pm 1.206}$	$6.091_{\pm 0.701}$	$6.182_{\pm 0.982}$	$\mathbf{2.727}_{\pm 1.348}$	$3.273_{\pm 0.467}$	$\underline{2.364}_{\pm 0.674}$
TVAE	$6.000_{\pm 0.775}$	$4.909_{\pm 0.539}$	$6.364_{\pm 0.505}$	$5.273_{\pm 1.104}$	$5.364_{\pm 1.433}$	$5.455_{\pm 1.036}$	$3.000_{\pm 1.414}$	$\underline{2.091}_{\pm 0.302}$	$\mathbf{1.364}_{\pm 0.674}$
TabDDPM	$3.889_{\pm 1.691}$	$4.000_{\pm 1.323}$	$3.667_{\pm 1.658}$	$3.889_{\pm 1.764}$	$3.333_{\pm 2.062}$	$4.333_{\pm 1.732}$	$3.222_{\pm 1.787}$	$4.778_{\pm 0.441}$	$7.000_{\pm 0.000}$
TabSyn	$\underline{2.364}_{\pm 0.809}$	$\underline{1.818}_{\pm 0.751}$	$\underline{2.364}_{\pm 0.924}$	$\underline{2.364}_{\pm 1.286}$	$2.727_{\pm 0.905}$	$3.000_{\pm 1.549}$	$\underline{2.818}_{\pm 1.779}$	$5.818_{\pm 0.405}$	$4.909_{\pm 0.701}$

Table 2: Big Benchmark: metric ranks

(\mu\pm\sigma)

across the 11 datasets of the Big Benchmark. Best results per metric are in bold , second best underlined .

Figure 1: Training times in minutes of XGenB-DF (v-DDIM specification) and XGenB-AR in the Small and Big Benchmark, respectively.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task	N	No. of features		No. of categories
Dataset	Task	N	categorical	numerical	min	max
acsincome ( Ding et al., 2021 )	Regression	1 664 500	8	3	2	529
adult ( Becker and Kohavi, 1996 )	Classification	48 842	9	6	2	42
bank ( Moro et al., 2014 )	Classification	41 188	11	10	2	12
beijing ( Liang et al., 2015 )	Regression	41 757	1	11	4	4
churn ( Keramati et al., 2014 )	Classification	3 150	5	9	2	5

Appendix

Table 4: Big Benchmark: overview of datasets.

Figure 2: Violinplot of metric scores for various levels of dropout in XGenB-DF, evaluated for 20 sampling seeds per dataset in the Small Benchmark. MLE uses ROCAUC and

R^{2}

for classification and regression, respectively.

dataset	acsincome		adult		diabetes		lending
model	Clustering	Naive	Clustering	Naive	Clustering	Naive	Clustering	Naive
metric
Shape	$0.987_{\pm 0.001}$	$0.987_{\pm 0.001}$	$0.994_{\pm 0.001}$	$0.994_{\pm 0.001}$	$0.995_{\pm 0.000}$	$0.995_{\pm 0.000}$	$0.984_{\pm 0.001}$	$0.984_{\pm 0.001}$
Trend	$0.946_{\pm 0.002}$	$0.942_{\pm 0.002}$	$0.980_{\pm 0.003}$	$0.980_{\pm 0.003}$	$0.978_{\pm 0.002}$	$0.977_{\pm 0.002}$	$0.952_{\pm 0.002}$	$0.952_{\pm 0.002}$
Detection	$0.689_{\pm 0.003}$	$0.764_{\pm 0.003}$	$0.614_{\pm 0.004}$	$0.614_{\pm 0.004}$	$0.774_{\pm 0.002}$	$0.781_{\pm 0.003}$	$0.652_{\pm 0.015}$	$0.660_{\pm 0.012}$
$\alpha$ -Precision	$0.992_{\pm 0.003}$	$0.993_{\pm 0.002}$	$0.997_{\pm 0.001}$	$0.996_{\pm 0.001}$	$0.996_{\pm 0.001}$	$0.996_{\pm 0.001}$	$0.972_{\pm 0.007}$	$0.977_{\pm 0.007}$

Appendix

Table 5: Ablation: effect of merging infrequent categories in high-cardinality categorical features through a clustering approach versus a naive approach - clustering all infrequent categories into a single category. Metric scores

(\mu\pm\sigma)

over 20 sampling seeds for datasets containing high-cardinality categorical features

(K>32)

. MLE uses ROCAUC and

R^{2}

for classification and regression, respectively.

metric	Shape	Trend	Detection	$\alpha$ -Precision	$\beta$ -Recall	MLE	DCR	Training Time	Inference Time
Win	5 (0.08%)	9 (0.27%)	9 (6.07%)	4 (0.59%)	10 (11.93%)	6 (8.30%)	3 (0.10%)	3 (29.59%)	1 (6.67%)
Loss	6 (0.07%)	2 (0.01%)	2 (1.00%)	7 (0.85%)	1 (1.68%)	5 (0.60%)	8 (0.46%)	8 (89.98%)	10 (80.72%)

Appendix

Table 6: Ablation: comparison between hierarchical classification and a multi-class XGBoost classifier for numerical features. Shows wins and losses of the hierarchical approach versus the multi-class approach over the 11 datasets of the Big Benchmark. Also shows corresponding average improvement/reduction in metric scores.

Figure 3: Violinplot of metric scores for various quantization strategies (Quantile, Uniform, KMeans) and de-quantization strategies (Uniform-sampling, EQF-sampling) in XGenBoost’s autoregressive model, evaluated for 20 sampling seeds per dataset in the Big Benchmark. MLE shows ROCAUC and

R^{2}

for classification and regression datasets, respectively.

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.927_{\pm 0.003}$	$0.911_{\pm 0.003}$	$0.949_{\pm 0.007}$	$0.908_{\pm 0.005}$	$0.951_{\pm 0.004}$	$0.714_{\pm 0.004}$	$0.798_{\pm 0.004}$	$0.957_{\pm 0.006}$	$\mathbf{0.974}_{\pm 0.004}$	$0.967_{\pm 0.005}$	$0.968_{\pm 0.005}$	$0.967_{\pm 0.005}$	$\underline{0.972}_{\pm 0.004}$
bean	$0.980_{\pm 0.001}$	$0.984_{\pm 0.002}$	$0.979_{\pm 0.001}$	$0.975_{\pm 0.001}$	$0.986_{\pm 0.002}$	$0.864_{\pm 0.002}$	$0.923_{\pm 0.002}$	$\mathbf{0.989}_{\pm 0.002}$	$0.987_{\pm 0.002}$	$0.987_{\pm 0.002}$	$\underline{0.988}_{\pm 0.002}$	$0.969_{\pm 0.002}$	$0.969_{\pm 0.002}$
blood transfusion	$0.957_{\pm 0.006}$	$0.949_{\pm 0.006}$	$0.954_{\pm 0.010}$	$0.926_{\pm 0.010}$	$0.942_{\pm 0.007}$	$0.791_{\pm 0.005}$	$0.824_{\pm 0.008}$	$0.955_{\pm 0.007}$	$\mathbf{0.968}_{\pm 0.007}$	$0.961_{\pm 0.010}$	$0.963_{\pm 0.008}$	$0.965_{\pm 0.010}$	$\underline{0.967}_{\pm 0.010}$
breast cancer diagnostic	$0.936_{\pm 0.004}$	$0.937_{\pm 0.005}$	$0.928_{\pm 0.010}$	$0.923_{\pm 0.009}$	$0.935_{\pm 0.007}$	$0.740_{\pm 0.003}$	$0.853_{\pm 0.006}$	$0.343_{\pm 0.007}$	$\mathbf{0.950}_{\pm 0.007}$	$0.945_{\pm 0.007}$	$\underline{0.947}_{\pm 0.007}$	$0.945_{\pm 0.006}$	$0.943_{\pm 0.005}$
california	$0.955_{\pm 0.001}$	$0.962_{\pm 0.001}$	$0.981_{\pm 0.001}$	$0.877_{\pm 0.001}$	$0.950_{\pm 0.001}$	$0.925_{\pm 0.001}$	$0.927_{\pm 0.001}$	$0.977_{\pm 0.001}$	$\mathbf{0.991}_{\pm 0.001}$	$0.981_{\pm 0.001}$	$0.982_{\pm 0.001}$	$0.985_{\pm 0.002}$	$\underline{0.988}_{\pm 0.001}$
car	$0.979_{\pm 0.002}$	$\underline{0.982}_{\pm 0.003}$	$0.979_{\pm 0.005}$	$0.981_{\pm 0.003}$	$0.979_{\pm 0.005}$	$0.895_{\pm 0.004}$	$0.906_{\pm 0.007}$	$0.980_{\pm 0.004}$	$\mathbf{0.983}_{\pm 0.003}$	$0.981_{\pm 0.004}$	$0.981_{\pm 0.003}$	$0.981_{\pm 0.004}$	$0.981_{\pm 0.003}$

Appendix

Table 7: Small Benchmark: Shape metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.985_{\pm 0.003}$	$0.983_{\pm 0.003}$	$0.976_{\pm 0.003}$	$0.987_{\pm 0.004}$	$\underline{0.988}_{\pm 0.002}$	$0.881_{\pm 0.004}$	$0.962_{\pm 0.005}$	$\underline{0.988}_{\pm 0.002}$	$0.986_{\pm 0.003}$	$\underline{0.988}_{\pm 0.003}$	$\mathbf{0.989}_{\pm 0.004}$	$\underline{0.988}_{\pm 0.004}$	$0.987_{\pm 0.004}$
bean	$0.982_{\pm 0.002}$	$0.924_{\pm 0.005}$	$0.983_{\pm 0.002}$	$0.986_{\pm 0.002}$	$\mathbf{0.990}_{\pm 0.002}$	$0.917_{\pm 0.001}$	$0.917_{\pm 0.002}$	$0.983_{\pm 0.002}$	$0.984_{\pm 0.002}$	$\underline{0.989}_{\pm 0.002}$	$\underline{0.989}_{\pm 0.002}$	$0.981_{\pm 0.002}$	$0.981_{\pm 0.002}$
blood transfusion	$0.938_{\pm 0.005}$	$0.870_{\pm 0.039}$	$0.943_{\pm 0.018}$	$\mathbf{0.965}_{\pm 0.005}$	$\underline{0.963}_{\pm 0.016}$	$0.668_{\pm 0.020}$	$0.742_{\pm 0.014}$	$0.962_{\pm 0.010}$	$0.955_{\pm 0.008}$	$0.961_{\pm 0.012}$	$0.960_{\pm 0.014}$	$0.962_{\pm 0.013}$	$0.961_{\pm 0.014}$
breast cancer diagnostic	$0.964_{\pm 0.004}$	$0.910_{\pm 0.004}$	$0.956_{\pm 0.004}$	$0.956_{\pm 0.006}$	$0.960_{\pm 0.006}$	$0.787_{\pm 0.002}$	$0.891_{\pm 0.005}$	$0.759_{\pm 0.002}$	$0.964_{\pm 0.003}$	$0.954_{\pm 0.006}$	$0.955_{\pm 0.006}$	$\underline{0.969}_{\pm 0.004}$	$\mathbf{0.971}_{\pm 0.004}$
california	$\underline{0.993}_{\pm 0.001}$	$0.988_{\pm 0.001}$	$0.988_{\pm 0.004}$	$0.987_{\pm 0.001}$	$\mathbf{0.994}_{\pm 0.001}$	$0.941_{\pm 0.001}$	$0.962_{\pm 0.001}$	$0.984_{\pm 0.003}$	$0.982_{\pm 0.003}$	$0.988_{\pm 0.006}$	$0.989_{\pm 0.005}$	$0.988_{\pm 0.005}$	$\underline{0.993}_{\pm 0.001}$
car	$0.950_{\pm 0.002}$	$0.960_{\pm 0.003}$	$0.959_{\pm 0.005}$	$0.959_{\pm 0.004}$	$0.958_{\pm 0.005}$	$0.842_{\pm 0.006}$	$0.823_{\pm 0.006}$	$0.938_{\pm 0.004}$	$\mathbf{0.962}_{\pm 0.003}$	$0.960_{\pm 0.004}$	$0.960_{\pm 0.004}$	$\underline{0.961}_{\pm 0.004}$	$\underline{0.961}_{\pm 0.003}$

Appendix

Table 8: Small Benchmark: Trend metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.985_{\pm 0.004}$	$0.996_{\pm 0.003}$	$0.951_{\pm 0.012}$	$0.991_{\pm 0.004}$	$0.982_{\pm 0.006}$	$1.000_{\pm 0.000}$	$0.998_{\pm 0.002}$	$0.929_{\pm 0.010}$	$0.712_{\pm 0.022}$	$0.652_{\pm 0.027}$	$\underline{0.582}_{\pm 0.024}$	$0.638_{\pm 0.022}$	$\mathbf{0.578}_{\pm 0.034}$
bean	$0.682_{\pm 0.007}$	$0.970_{\pm 0.006}$	$0.933_{\pm 0.008}$	$0.990_{\pm 0.002}$	$0.625_{\pm 0.014}$	$1.000_{\pm 0.000}$	$1.000_{\pm 0.000}$	$\mathbf{0.511}_{\pm 0.007}$	$0.898_{\pm 0.014}$	$0.555_{\pm 0.008}$	$\underline{0.543}_{\pm 0.009}$	$0.595_{\pm 0.005}$	$0.604_{\pm 0.008}$
blood transfusion	$0.651_{\pm 0.030}$	$0.877_{\pm 0.014}$	$0.768_{\pm 0.022}$	$0.964_{\pm 0.009}$	$0.953_{\pm 0.013}$	$0.996_{\pm 0.002}$	$0.996_{\pm 0.002}$	$0.901_{\pm 0.014}$	$0.525_{\pm 0.043}$	$0.530_{\pm 0.033}$	$0.546_{\pm 0.034}$	$\mathbf{0.519}_{\pm 0.041}$	$\underline{0.521}_{\pm 0.038}$
breast cancer diagnostic	$0.598_{\pm 0.025}$	$0.929_{\pm 0.025}$	$0.614_{\pm 0.025}$	$0.704_{\pm 0.042}$	$0.613_{\pm 0.032}$	$1.000_{\pm 0.000}$	$0.952_{\pm 0.016}$	$1.000_{\pm 0.000}$	$\underline{0.565}_{\pm 0.037}$	$0.658_{\pm 0.030}$	$0.590_{\pm 0.038}$	$0.615_{\pm 0.036}$	$\mathbf{0.560}_{\pm 0.029}$
california	$0.805_{\pm 0.005}$	$0.875_{\pm 0.004}$	$0.721_{\pm 0.005}$	$0.976_{\pm 0.001}$	$0.816_{\pm 0.006}$	$0.969_{\pm 0.001}$	$0.935_{\pm 0.002}$	$0.777_{\pm 0.006}$	$0.706_{\pm 0.008}$	$0.677_{\pm 0.006}$	$\underline{0.641}_{\pm 0.007}$	$0.665_{\pm 0.009}$	$\mathbf{0.630}_{\pm 0.008}$
car	$0.800_{\pm 0.012}$	$0.501_{\pm 0.016}$	$0.506_{\pm 0.021}$	$\underline{0.490}_{\pm 0.033}$	$\mathbf{0.488}_{\pm 0.022}$	$0.785_{\pm 0.016}$	$0.872_{\pm 0.008}$	$0.696_{\pm 0.022}$	$0.501_{\pm 0.003}$	$0.501_{\pm 0.030}$	$0.505_{\pm 0.021}$	$0.516_{\pm 0.026}$	$0.499_{\pm 0.008}$

Appendix

Table 9: Small Benchmark: Detection metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.891_{\pm 0.009}$	$0.966_{\pm 0.016}$	$0.972_{\pm 0.007}$	$0.958_{\pm 0.019}$	$0.953_{\pm 0.020}$	$0.692_{\pm 0.017}$	$0.933_{\pm 0.008}$	$0.971_{\pm 0.013}$	$\mathbf{0.980}_{\pm 0.007}$	$\underline{0.979}_{\pm 0.009}$	$0.955_{\pm 0.015}$	$0.973_{\pm 0.013}$	$0.975_{\pm 0.010}$
bean	$0.965_{\pm 0.003}$	$0.983_{\pm 0.004}$	$0.948_{\pm 0.004}$	$\underline{0.988}_{\pm 0.004}$	$0.987_{\pm 0.006}$	$0.746_{\pm 0.006}$	$0.928_{\pm 0.005}$	$\mathbf{0.992}_{\pm 0.003}$	$0.982_{\pm 0.005}$	$0.983_{\pm 0.005}$	$0.981_{\pm 0.005}$	$0.900_{\pm 0.004}$	$0.899_{\pm 0.004}$
blood transfusion	$0.944_{\pm 0.014}$	$0.961_{\pm 0.020}$	$0.955_{\pm 0.020}$	$0.955_{\pm 0.012}$	$0.965_{\pm 0.016}$	$0.611_{\pm 0.028}$	$0.671_{\pm 0.015}$	$\mathbf{0.971}_{\pm 0.013}$	$\underline{0.968}_{\pm 0.013}$	$0.960_{\pm 0.017}$	$0.956_{\pm 0.021}$	$0.963_{\pm 0.016}$	$0.960_{\pm 0.016}$
breast cancer diagnostic	$0.891_{\pm 0.017}$	$0.884_{\pm 0.022}$	$0.818_{\pm 0.031}$	$0.898_{\pm 0.040}$	$0.886_{\pm 0.034}$	$0.369_{\pm 0.010}$	$0.734_{\pm 0.017}$	$0.000_{\pm 0.000}$	$\underline{0.949}_{\pm 0.025}$	$\mathbf{0.965}_{\pm 0.011}$	$0.939_{\pm 0.018}$	$0.876_{\pm 0.029}$	$0.850_{\pm 0.031}$
california	$0.868_{\pm 0.003}$	$\underline{0.978}_{\pm 0.005}$	$0.973_{\pm 0.004}$	$0.915_{\pm 0.005}$	$0.945_{\pm 0.005}$	$0.933_{\pm 0.005}$	$0.944_{\pm 0.005}$	$0.959_{\pm 0.005}$	$\mathbf{0.984}_{\pm 0.004}$	$0.943_{\pm 0.004}$	$0.928_{\pm 0.004}$	$0.970_{\pm 0.007}$	$0.960_{\pm 0.007}$
car	$0.966_{\pm 0.009}$	$\mathbf{0.982}_{\pm 0.007}$	$0.976_{\pm 0.010}$	$\mathbf{0.982}_{\pm 0.005}$	$0.977_{\pm 0.009}$	$0.951_{\pm 0.008}$	$0.865_{\pm 0.016}$	$0.949_{\pm 0.015}$	$\underline{0.981}_{\pm 0.007}$	$0.980_{\pm 0.006}$	$\mathbf{0.982}_{\pm 0.007}$	$\underline{0.981}_{\pm 0.009}$	$\mathbf{0.982}_{\pm 0.007}$

Appendix

Table 10: Small Benchmark:

\alpha

-Precision metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.299_{\pm 0.012}$	$0.067_{\pm 0.009}$	$0.416_{\pm 0.018}$	$0.143_{\pm 0.012}$	$\mathbf{0.606}_{\pm 0.022}$	$0.002_{\pm 0.002}$	$0.073_{\pm 0.008}$	$\underline{0.586}_{\pm 0.017}$	$0.468_{\pm 0.017}$	$0.511_{\pm 0.017}$	$0.549_{\pm 0.020}$	$0.490_{\pm 0.015}$	$0.519_{\pm 0.016}$
bean	$\mathbf{0.668}_{\pm 0.003}$	$0.134_{\pm 0.003}$	$0.200_{\pm 0.004}$	$0.062_{\pm 0.002}$	$0.565_{\pm 0.008}$	$0.000_{\pm 0.000}$	$0.003_{\pm 0.000}$	$0.498_{\pm 0.008}$	$0.271_{\pm 0.006}$	$\underline{0.573}_{\pm 0.005}$	$0.562_{\pm 0.005}$	$0.564_{\pm 0.006}$	$0.547_{\pm 0.007}$
blood transfusion	$\underline{0.771}_{\pm 0.013}$	$0.436_{\pm 0.028}$	$0.733_{\pm 0.035}$	$0.336_{\pm 0.025}$	$0.610_{\pm 0.026}$	$0.084_{\pm 0.022}$	$0.140_{\pm 0.019}$	$0.522_{\pm 0.028}$	$0.692_{\pm 0.025}$	$0.764_{\pm 0.029}$	$\mathbf{0.779}_{\pm 0.017}$	$0.756_{\pm 0.027}$	$0.768_{\pm 0.029}$
breast cancer diagnostic	$0.821_{\pm 0.013}$	$0.051_{\pm 0.011}$	$0.577_{\pm 0.029}$	$0.542_{\pm 0.024}$	$\underline{0.828}_{\pm 0.022}$	$0.000_{\pm 0.000}$	$0.023_{\pm 0.008}$	$0.000_{\pm 0.000}$	$0.583_{\pm 0.021}$	$\mathbf{0.835}_{\pm 0.022}$	$0.809_{\pm 0.017}$	$0.826_{\pm 0.018}$	$0.771_{\pm 0.027}$
california	$\mathbf{0.548}_{\pm 0.004}$	$0.252_{\pm 0.005}$	$0.420_{\pm 0.004}$	$0.103_{\pm 0.003}$	$\underline{0.509}_{\pm 0.004}$	$0.164_{\pm 0.003}$	$0.273_{\pm 0.003}$	$0.395_{\pm 0.005}$	$0.380_{\pm 0.004}$	$0.434_{\pm 0.005}$	$0.445_{\pm 0.006}$	$0.415_{\pm 0.005}$	$0.425_{\pm 0.005}$
car	$\underline{0.987}_{\pm 0.003}$	$\mathbf{0.995}_{\pm 0.002}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$	$0.959_{\pm 0.005}$	$0.927_{\pm 0.008}$	$0.980_{\pm 0.004}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$	$\mathbf{0.995}_{\pm 0.001}$

Appendix

Table 11: Small Benchmark:

\beta

-Recall metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	Real	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.922_{\pm 0.000}$	$0.837_{\pm 0.018}$	$0.634_{\pm 0.041}$	$0.736_{\pm 0.032}$	$0.741_{\pm 0.024}$	$\underline{0.845}_{\pm 0.023}$	$-1.163_{\pm 0.318}$	$0.115_{\pm 0.135}$	$\mathbf{0.878}_{\pm 0.013}$	$0.737_{\pm 0.029}$	$0.830_{\pm 0.017}$	$0.817_{\pm 0.016}$	$0.827_{\pm 0.017}$	$0.812_{\pm 0.026}$
california	$0.826_{\pm 0.003}$	$0.761_{\pm 0.004}$	$0.680_{\pm 0.009}$	$0.762_{\pm 0.004}$	$0.688_{\pm 0.009}$	$0.711_{\pm 0.006}$	$0.389_{\pm 0.017}$	$0.662_{\pm 0.008}$	$0.768_{\pm 0.005}$	$0.727_{\pm 0.006}$	$\mathbf{0.772}_{\pm 0.004}$	$\underline{0.770}_{\pm 0.004}$	$0.768_{\pm 0.004}$	$0.767_{\pm 0.004}$
concrete compression	$0.913_{\pm 0.000}$	$\mathbf{0.851}_{\pm 0.014}$	$0.452_{\pm 0.073}$	$0.704_{\pm 0.034}$	$0.744_{\pm 0.021}$	$0.813_{\pm 0.018}$	$-0.776_{\pm 0.255}$	$0.435_{\pm 0.109}$	$\underline{0.835}_{\pm 0.016}$	$0.749_{\pm 0.025}$	$0.818_{\pm 0.018}$	$0.810_{\pm 0.019}$	$0.806_{\pm 0.025}$	$0.790_{\pm 0.028}$
concrete slump	$0.837_{\pm 0.000}$	$\mathbf{0.774}_{\pm 0.041}$	$0.034_{\pm 0.300}$	$0.515_{\pm 0.114}$	$\underline{0.742}_{\pm 0.075}$	$0.723_{\pm 0.092}$	$-1.049_{\pm 0.615}$	$0.044_{\pm 0.266}$	$-0.768_{\pm 0.935}$	$0.595_{\pm 0.112}$	$0.667_{\pm 0.120}$	$0.627_{\pm 0.095}$	$0.616_{\pm 0.144}$	$0.627_{\pm 0.122}$
wine quality red	$0.446_{\pm 0.000}$	$\mathbf{0.387}_{\pm 0.025}$	$0.260_{\pm 0.043}$	$0.301_{\pm 0.043}$	$0.324_{\pm 0.035}$	$0.326_{\pm 0.034}$	$-0.066_{\pm 0.062}$	$0.301_{\pm 0.015}$	$0.329_{\pm 0.040}$	$0.292_{\pm 0.033}$	$0.347_{\pm 0.030}$	$\underline{0.348}_{\pm 0.037}$	$0.328_{\pm 0.037}$	$0.332_{\pm 0.031}$
wine quality white	$0.486_{\pm 0.000}$	$\mathbf{0.336}_{\pm 0.023}$	$0.171_{\pm 0.020}$	$0.213_{\pm 0.026}$	$0.223_{\pm 0.029}$	$0.235_{\pm 0.021}$	$-0.316_{\pm 0.060}$	$0.223_{\pm 0.023}$	$\underline{0.254}_{\pm 0.017}$	$0.189_{\pm 0.028}$	$0.253_{\pm 0.025}$	$0.250_{\pm 0.019}$	$0.243_{\pm 0.029}$	$0.234_{\pm 0.032}$

Appendix

Table 12: Small Benchmark: MLE (

R^{2}

) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
bean	$0.997_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\underline{0.995}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$0.983_{\pm 0.001}$	$0.989_{\pm 0.001}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$	$\mathbf{0.996}_{\pm 0.000}$
blood transfusion	$0.783_{\pm 0.010}$	$\mathbf{0.780}_{\pm 0.010}$	$0.766_{\pm 0.033}$	$0.746_{\pm 0.020}$	$\underline{0.775}_{\pm 0.019}$	$0.761_{\pm 0.023}$	$0.561_{\pm 0.090}$	$0.500_{\pm 0.000}$	$0.728_{\pm 0.025}$	$0.727_{\pm 0.030}$	$0.743_{\pm 0.025}$	$0.741_{\pm 0.030}$	$0.731_{\pm 0.029}$	$0.728_{\pm 0.026}$
breast cancer diagnostic	$0.991_{\pm 0.000}$	$\underline{0.990}_{\pm 0.002}$	$0.982_{\pm 0.009}$	$0.984_{\pm 0.006}$	$\underline{0.990}_{\pm 0.002}$	$0.989_{\pm 0.003}$	$0.848_{\pm 0.078}$	$0.988_{\pm 0.004}$	$0.499_{\pm 0.050}$	$\underline{0.990}_{\pm 0.003}$	$\mathbf{0.991}_{\pm 0.003}$	$\mathbf{0.991}_{\pm 0.002}$	$\mathbf{0.991}_{\pm 0.003}$	$\mathbf{0.991}_{\pm 0.002}$
car	$1.000_{\pm 0.000}$	$0.993_{\pm 0.001}$	$0.985_{\pm 0.003}$	$0.996_{\pm 0.001}$	$0.987_{\pm 0.002}$	$0.988_{\pm 0.002}$	$0.855_{\pm 0.014}$	$0.950_{\pm 0.004}$	$0.815_{\pm 0.013}$	$0.986_{\pm 0.004}$	$\mathbf{0.998}_{\pm 0.001}$	$\mathbf{0.998}_{\pm 0.001}$	$\underline{0.997}_{\pm 0.001}$	$\mathbf{0.998}_{\pm 0.001}$
climate model crashes	$0.919_{\pm 0.005}$	$0.865_{\pm 0.021}$	$0.720_{\pm 0.093}$	$0.863_{\pm 0.049}$	$\mathbf{0.894}_{\pm 0.029}$	$0.885_{\pm 0.026}$	$0.570_{\pm 0.069}$	$0.500_{\pm 0.000}$	$0.881_{\pm 0.032}$	$\underline{0.889}_{\pm 0.028}$	$0.839_{\pm 0.026}$	$0.868_{\pm 0.031}$	$0.862_{\pm 0.031}$	$0.868_{\pm 0.053}$
congress	$0.988_{\pm 0.000}$	$\underline{0.986}_{\pm 0.004}$	$0.968_{\pm 0.013}$	$0.981_{\pm 0.012}$	$0.977_{\pm 0.006}$	$0.977_{\pm 0.009}$	$0.558_{\pm 0.192}$	$0.962_{\pm 0.014}$	$\mathbf{0.989}_{\pm 0.005}$	$0.982_{\pm 0.010}$	$0.970_{\pm 0.011}$	$0.973_{\pm 0.011}$	$0.973_{\pm 0.013}$	$0.969_{\pm 0.017}$

Appendix

Table 13: Small Benchmark: MLE (ROCAUC) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
bean	$0.931_{\pm 0.001}$	$0.927_{\pm 0.003}$	$0.915_{\pm 0.003}$	$0.924_{\pm 0.002}$	$0.923_{\pm 0.001}$	$\underline{0.928}_{\pm 0.003}$	$0.858_{\pm 0.006}$	$0.890_{\pm 0.001}$	$\mathbf{0.929}_{\pm 0.003}$	$0.924_{\pm 0.002}$	$0.926_{\pm 0.002}$	$0.925_{\pm 0.002}$	$0.925_{\pm 0.002}$	$0.925_{\pm 0.003}$
blood transfusion	$0.432_{\pm 0.044}$	$\mathbf{0.475}_{\pm 0.046}$	$0.344_{\pm 0.138}$	$0.421_{\pm 0.059}$	$0.371_{\pm 0.111}$	$0.413_{\pm 0.137}$	$0.285_{\pm 0.115}$	$0.000_{\pm 0.000}$	$\underline{0.438}_{\pm 0.081}$	$0.372_{\pm 0.116}$	$0.432_{\pm 0.080}$	$0.412_{\pm 0.083}$	$0.366_{\pm 0.124}$	$0.370_{\pm 0.119}$
breast cancer diagnostic	$0.947_{\pm 0.004}$	$0.943_{\pm 0.009}$	$0.950_{\pm 0.016}$	$0.948_{\pm 0.009}$	$0.949_{\pm 0.008}$	$0.951_{\pm 0.009}$	$0.771_{\pm 0.091}$	$0.932_{\pm 0.015}$	$0.582_{\pm 0.342}$	$\underline{0.956}_{\pm 0.011}$	$\mathbf{0.957}_{\pm 0.010}$	$0.954_{\pm 0.009}$	$0.950_{\pm 0.010}$	$0.951_{\pm 0.012}$
car	$0.977_{\pm 0.004}$	$0.915_{\pm 0.013}$	$0.874_{\pm 0.013}$	$0.944_{\pm 0.011}$	$0.884_{\pm 0.018}$	$0.891_{\pm 0.017}$	$0.648_{\pm 0.026}$	$0.760_{\pm 0.016}$	$0.585_{\pm 0.026}$	$0.883_{\pm 0.020}$	$\mathbf{0.961}_{\pm 0.012}$	$0.955_{\pm 0.012}$	$0.953_{\pm 0.010}$	$\underline{0.960}_{\pm 0.011}$
climate model crashes	$0.975_{\pm 0.005}$	$\mathbf{0.975}_{\pm 0.006}$	$0.956_{\pm 0.003}$	$0.957_{\pm 0.001}$	$0.972_{\pm 0.007}$	$0.972_{\pm 0.006}$	$0.892_{\pm 0.046}$	$0.957_{\pm 0.000}$	$\underline{0.973}_{\pm 0.006}$	$0.963_{\pm 0.007}$	$0.966_{\pm 0.009}$	$0.969_{\pm 0.007}$	$0.967_{\pm 0.008}$	$0.970_{\pm 0.007}$
congress	$0.947_{\pm 0.005}$	$\underline{0.950}_{\pm 0.004}$	$0.925_{\pm 0.014}$	$0.943_{\pm 0.008}$	$0.943_{\pm 0.014}$	$0.938_{\pm 0.007}$	$0.605_{\pm 0.127}$	$0.915_{\pm 0.015}$	$\mathbf{0.952}_{\pm 0.008}$	$0.945_{\pm 0.006}$	$0.946_{\pm 0.006}$	$0.947_{\pm 0.009}$	$0.943_{\pm 0.011}$	$0.946_{\pm 0.007}$

Appendix

Table 14: Small Benchmark: MLE (F1) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.291_{\pm 0.000}$	$0.421_{\pm 0.023}$	$0.629_{\pm 0.035}$	$0.535_{\pm 0.033}$	$0.530_{\pm 0.025}$	$\underline{0.410}_{\pm 0.030}$	$1.528_{\pm 0.110}$	$0.977_{\pm 0.073}$	$\mathbf{0.363}_{\pm 0.019}$	$0.533_{\pm 0.029}$	$0.429_{\pm 0.021}$	$0.446_{\pm 0.019}$	$0.433_{\pm 0.021}$	$0.451_{\pm 0.031}$
california	$0.415_{\pm 0.004}$	$0.486_{\pm 0.004}$	$0.562_{\pm 0.008}$	$0.485_{\pm 0.004}$	$0.555_{\pm 0.008}$	$0.534_{\pm 0.006}$	$0.778_{\pm 0.011}$	$0.578_{\pm 0.007}$	$0.480_{\pm 0.005}$	$0.519_{\pm 0.006}$	$\mathbf{0.475}_{\pm 0.004}$	$\underline{0.477}_{\pm 0.004}$	$0.479_{\pm 0.004}$	$0.481_{\pm 0.004}$
concrete compression	$0.285_{\pm 0.000}$	$\mathbf{0.371}_{\pm 0.017}$	$0.712_{\pm 0.047}$	$0.524_{\pm 0.030}$	$0.488_{\pm 0.020}$	$0.417_{\pm 0.020}$	$1.282_{\pm 0.093}$	$0.722_{\pm 0.067}$	$\underline{0.391}_{\pm 0.018}$	$0.482_{\pm 0.024}$	$0.411_{\pm 0.021}$	$0.420_{\pm 0.021}$	$0.424_{\pm 0.028}$	$0.441_{\pm 0.030}$
concrete slump	$0.376_{\pm 0.000}$	$\mathbf{0.442}_{\pm 0.040}$	$0.907_{\pm 0.142}$	$0.646_{\pm 0.077}$	$\underline{0.469}_{\pm 0.071}$	$0.485_{\pm 0.077}$	$1.321_{\pm 0.204}$	$0.904_{\pm 0.129}$	$1.206_{\pm 0.302}$	$0.589_{\pm 0.079}$	$0.530_{\pm 0.098}$	$0.566_{\pm 0.073}$	$0.569_{\pm 0.109}$	$0.563_{\pm 0.091}$
wine quality red	$0.744_{\pm 0.000}$	$\mathbf{0.783}_{\pm 0.016}$	$0.860_{\pm 0.025}$	$0.836_{\pm 0.026}$	$0.822_{\pm 0.022}$	$0.821_{\pm 0.021}$	$1.032_{\pm 0.030}$	$0.836_{\pm 0.009}$	$0.819_{\pm 0.024}$	$0.841_{\pm 0.020}$	$0.808_{\pm 0.018}$	$\underline{0.807}_{\pm 0.023}$	$0.820_{\pm 0.022}$	$0.817_{\pm 0.019}$
wine quality white	$0.706_{\pm 0.000}$	$\mathbf{0.802}_{\pm 0.014}$	$0.897_{\pm 0.011}$	$0.874_{\pm 0.014}$	$0.868_{\pm 0.016}$	$0.861_{\pm 0.012}$	$1.129_{\pm 0.026}$	$0.868_{\pm 0.013}$	$\underline{0.851}_{\pm 0.010}$	$0.887_{\pm 0.015}$	$\underline{0.851}_{\pm 0.014}$	$0.853_{\pm 0.011}$	$0.857_{\pm 0.016}$	$0.862_{\pm 0.018}$

Appendix

Table 15: Small Benchmark: MLE (RMSE) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.722_{\pm 0.039}$	$\underline{0.916}_{\pm 0.030}$	$0.864_{\pm 0.047}$	$0.857_{\pm 0.039}$	$0.836_{\pm 0.026}$	$\mathbf{0.939}_{\pm 0.042}$	$0.901_{\pm 0.038}$	$0.818_{\pm 0.048}$	$0.878_{\pm 0.043}$	$0.852_{\pm 0.036}$	$0.828_{\pm 0.040}$	$0.859_{\pm 0.050}$	$0.847_{\pm 0.048}$
bean	$0.817_{\pm 0.010}$	$0.915_{\pm 0.014}$	$0.934_{\pm 0.015}$	$0.962_{\pm 0.014}$	$0.905_{\pm 0.015}$	$\mathbf{0.987}_{\pm 0.011}$	$0.967_{\pm 0.010}$	$\underline{0.976}_{\pm 0.011}$	$0.943_{\pm 0.013}$	$0.901_{\pm 0.015}$	$0.907_{\pm 0.013}$	$0.917_{\pm 0.015}$	$0.922_{\pm 0.015}$
blood transfusion	$0.988_{\pm 0.019}$	$\mathbf{1.000}_{\pm 0.000}$	$0.965_{\pm 0.032}$	$0.983_{\pm 0.023}$	$0.986_{\pm 0.024}$	$0.984_{\pm 0.017}$	$0.983_{\pm 0.030}$	$\underline{0.996}_{\pm 0.009}$	$\underline{0.996}_{\pm 0.010}$	$0.990_{\pm 0.018}$	$0.991_{\pm 0.012}$	$\underline{0.996}_{\pm 0.011}$	$0.988_{\pm 0.014}$
breast cancer diagnostic	$0.667_{\pm 0.055}$	$\underline{0.892}_{\pm 0.054}$	$0.718_{\pm 0.069}$	$0.741_{\pm 0.068}$	$0.690_{\pm 0.076}$	$\mathbf{0.947}_{\pm 0.048}$	$0.847_{\pm 0.070}$	$0.723_{\pm 0.083}$	$0.718_{\pm 0.071}$	$0.644_{\pm 0.066}$	$0.661_{\pm 0.068}$	$0.692_{\pm 0.057}$	$0.704_{\pm 0.058}$
california	$0.842_{\pm 0.012}$	$0.968_{\pm 0.011}$	$0.939_{\pm 0.010}$	$0.964_{\pm 0.007}$	$0.941_{\pm 0.011}$	$\underline{0.974}_{\pm 0.011}$	$\mathbf{0.986}_{\pm 0.010}$	$0.938_{\pm 0.012}$	$0.962_{\pm 0.010}$	$0.964_{\pm 0.010}$	$0.961_{\pm 0.013}$	$0.966_{\pm 0.008}$	$0.962_{\pm 0.010}$
car	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$	$\mathbf{1.000}_{\pm 0.000}$

Appendix

Table 16: Small Benchmark: DCR metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	-	$\underline{0.050}$	$0.205$	$0.346$	$0.371$	$0.067$	$0.081$	$1.158$	$\mathbf{0.018}$	$0.339$	$0.326$	$0.419$	$0.298$
bean	-	$4.083$	$9.696$	$16.384$	$16.676$	$1.940$	$\underline{1.208}$	$9.132$	$\mathbf{0.069}$	$18.009$	$15.328$	$13.069$	$14.507$
blood transfusion	-	$\underline{0.031}$	$0.120$	$0.143$	$0.137$	$0.080$	$0.038$	$0.729$	$\mathbf{0.011}$	$0.134$	$0.140$	$0.130$	$0.134$
breast cancer diagnostic	-	$0.178$	$2.439$	$6.791$	$6.375$	$0.211$	$\underline{0.151}$	$0.674$	$\mathbf{0.102}$	$6.213$	$6.016$	$5.511$	$5.755$
california	-	$3.100$	$3.789$	$9.872$	$10.087$	$1.695$	$\underline{1.264}$	$14.155$	$\mathbf{0.037}$	$14.022$	$10.392$	$8.443$	$10.041$
car	-	$0.069$	$0.123$	$3.478$	$3.448$	$0.154$	$\underline{0.065}$	$1.814$	$\mathbf{0.001}$	$2.858$	$2.964$	$2.773$	$2.916$

Appendix

Table 17: Small Benchmark: training time in minutes for all datasets. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	UT	FD	FF	CTGAN	TVAE	TabDDPM	XGenB-AR	XGenB-DF (x-DDPM)	XGenB-DF (v-DDPM)	XGenB-DF (x-DDIM)	XGenB-DF (v-DDIM)
airfoil self noise	$0.039_{\pm 0.000}$	$\underline{0.035}_{\pm 0.000}$	$17.895_{\pm 1.361}$	$0.322_{\pm 0.005}$	$0.153_{\pm 0.001}$	$\mathbf{0.023}_{\pm 0.000}$	$\mathbf{0.023}_{\pm 0.000}$	$1.356_{\pm 0.010}$	$0.153_{\pm 0.000}$	$0.723_{\pm 0.003}$	$0.709_{\pm 0.003}$	$0.692_{\pm 0.005}$	$0.709_{\pm 0.010}$
bean	$0.318_{\pm 0.011}$	$1.713_{\pm 0.020}$	$574.660_{\pm 4.101}$	$5.602_{\pm 0.096}$	$2.596_{\pm 0.047}$	$\underline{0.142}_{\pm 0.001}$	$\mathbf{0.095}_{\pm 0.000}$	$15.054_{\pm 0.058}$	$0.907_{\pm 0.003}$	$8.309_{\pm 0.106}$	$8.613_{\pm 0.053}$	$8.137_{\pm 0.221}$	$8.070_{\pm 0.108}$
blood transfusion	$0.040_{\pm 0.000}$	$0.086_{\pm 0.001}$	$6.601_{\pm 0.070}$	$0.300_{\pm 0.005}$	$0.154_{\pm 0.002}$	$\mathbf{0.019}_{\pm 0.000}$	$\underline{0.027}_{\pm 0.000}$	$2.556_{\pm 0.014}$	$0.092_{\pm 0.000}$	$1.376_{\pm 0.008}$	$1.363_{\pm 0.006}$	$1.372_{\pm 0.006}$	$1.358_{\pm 0.005}$
breast cancer diagnostic	$\mathbf{0.022}_{\pm 0.000}$	$0.219_{\pm 0.002}$	$44.544_{\pm 0.877}$	$1.493_{\pm 0.080}$	$0.733_{\pm 0.020}$	$\underline{0.081}_{\pm 0.014}$	$0.137_{\pm 0.001}$	$4.102_{\pm 0.011}$	$0.806_{\pm 0.004}$	$1.792_{\pm 0.037}$	$2.592_{\pm 0.024}$	$2.093_{\pm 0.135}$	$2.577_{\pm 0.017}$
california	$\underline{0.081}_{\pm 0.006}$	$0.342_{\pm 0.008}$	$483.389_{\pm 11.395}$	$3.181_{\pm 0.028}$	$1.389_{\pm 0.015}$	$0.119_{\pm 0.001}$	$\mathbf{0.076}_{\pm 0.001}$	$22.548_{\pm 0.400}$	$0.583_{\pm 0.005}$	$3.104_{\pm 0.031}$	$3.627_{\pm 0.100}$	$4.174_{\pm 0.136}$	$3.044_{\pm 0.052}$
car	$0.579_{\pm 0.006}$	$1.020_{\pm 0.020}$	$7.883_{\pm 0.062}$	$1.510_{\pm 0.034}$	$0.728_{\pm 0.016}$	$\underline{0.026}_{\pm 0.000}$	$\mathbf{0.022}_{\pm 0.000}$	$6.951_{\pm 0.021}$	$0.049_{\pm 0.000}$	$5.131_{\pm 0.322}$	$5.463_{\pm 0.290}$	$5.292_{\pm 0.377}$	$5.026_{\pm 0.329}$

Appendix

Table 18: Small Benchmark: inference time in seconds for all datasets. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$0.977_{\pm 0.001}$	$0.856_{\pm 0.001}$	$0.855_{\pm 0.001}$	-	$\underline{0.978}_{\pm 0.001}$	$\mathbf{0.987}_{\pm 0.001}$
adult	$0.925_{\pm 0.000}$	$0.956_{\pm 0.001}$	$0.842_{\pm 0.001}$	$0.913_{\pm 0.001}$	$0.987_{\pm 0.001}$	$\underline{0.990}_{\pm 0.000}$	$\mathbf{0.994}_{\pm 0.001}$
bank	$0.961_{\pm 0.000}$	$0.972_{\pm 0.001}$	$0.895_{\pm 0.000}$	$0.887_{\pm 0.001}$	$0.980_{\pm 0.001}$	$\underline{0.990}_{\pm 0.000}$	$\mathbf{0.992}_{\pm 0.001}$
beijing	$0.978_{\pm 0.000}$	$0.974_{\pm 0.001}$	$0.917_{\pm 0.001}$	$0.889_{\pm 0.001}$	$0.985_{\pm 0.001}$	$\underline{0.992}_{\pm 0.001}$	$\mathbf{0.993}_{\pm 0.001}$
churn	$0.972_{\pm 0.001}$	$0.966_{\pm 0.002}$	$0.836_{\pm 0.003}$	$0.938_{\pm 0.002}$	$0.922_{\pm 0.003}$	$\underline{0.975}_{\pm 0.003}$	$\mathbf{0.982}_{\pm 0.002}$
covertype	-	$0.995_{\pm 0.000}$	$0.955_{\pm 0.000}$	$0.980_{\pm 0.000}$	$\underline{0.996}_{\pm 0.000}$	$0.990_{\pm 0.000}$	$\mathbf{0.997}_{\pm 0.000}$

Appendix

Table 19: Big Benchmark: Shape metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$0.712_{\pm 0.018}$	$0.710_{\pm 0.003}$	$0.661_{\pm 0.010}$	-	$\underline{0.941}_{\pm 0.002}$	$\mathbf{0.946}_{\pm 0.002}$
adult	$0.859_{\pm 0.004}$	$0.636_{\pm 0.011}$	$0.782_{\pm 0.005}$	$0.838_{\pm 0.025}$	$0.963_{\pm 0.001}$	$\underline{0.977}_{\pm 0.001}$	$\mathbf{0.980}_{\pm 0.003}$
bank	$0.931_{\pm 0.003}$	$0.572_{\pm 0.010}$	$0.759_{\pm 0.006}$	$0.810_{\pm 0.003}$	$0.956_{\pm 0.001}$	$\mathbf{0.980}_{\pm 0.001}$	$\underline{0.978}_{\pm 0.002}$
beijing	$0.989_{\pm 0.001}$	$0.892_{\pm 0.003}$	$0.901_{\pm 0.001}$	$0.912_{\pm 0.003}$	$0.985_{\pm 0.001}$	$\mathbf{0.994}_{\pm 0.001}$	$\underline{0.990}_{\pm 0.001}$
churn	$\underline{0.961}_{\pm 0.002}$	$0.735_{\pm 0.018}$	$0.775_{\pm 0.007}$	$0.901_{\pm 0.004}$	$0.885_{\pm 0.004}$	$\underline{0.961}_{\pm 0.002}$	$\mathbf{0.965}_{\pm 0.004}$
covertype	-	$0.819_{\pm 0.013}$	$0.900_{\pm 0.004}$	$0.920_{\pm 0.008}$	$0.973_{\pm 0.003}$	$\underline{0.979}_{\pm 0.003}$	$\mathbf{0.988}_{\pm 0.004}$

Appendix

Table 20: Big Benchmark: Trend metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$0.842_{\pm 0.003}$	$0.964_{\pm 0.001}$	$0.974_{\pm 0.000}$	-	$\mathbf{0.611}_{\pm 0.003}$	$\underline{0.689}_{\pm 0.003}$
adult	$0.956_{\pm 0.001}$	$0.956_{\pm 0.002}$	$0.998_{\pm 0.000}$	$0.978_{\pm 0.001}$	$\underline{0.619}_{\pm 0.004}$	$0.668_{\pm 0.004}$	$\mathbf{0.614}_{\pm 0.004}$
bank	$0.917_{\pm 0.003}$	$0.927_{\pm 0.002}$	$1.000_{\pm 0.000}$	$0.998_{\pm 0.000}$	$0.881_{\pm 0.004}$	$\underline{0.866}_{\pm 0.003}$	$\mathbf{0.817}_{\pm 0.003}$
beijing	$0.938_{\pm 0.004}$	$0.985_{\pm 0.002}$	$0.997_{\pm 0.000}$	$0.998_{\pm 0.000}$	$0.876_{\pm 0.003}$	$\mathbf{0.768}_{\pm 0.007}$	$\underline{0.813}_{\pm 0.005}$
churn	$\underline{0.778}_{\pm 0.011}$	$0.900_{\pm 0.011}$	$0.999_{\pm 0.001}$	$0.977_{\pm 0.003}$	$0.932_{\pm 0.007}$	$0.866_{\pm 0.011}$	$\mathbf{0.739}_{\pm 0.016}$
covertype	-	$0.975_{\pm 0.001}$	$1.000_{\pm 0.000}$	$0.998_{\pm 0.000}$	$\mathbf{0.800}_{\pm 0.005}$	$\underline{0.844}_{\pm 0.003}$	$0.873_{\pm 0.003}$

Appendix

Table 21: Big Benchmark: Detection metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$\underline{0.991}_{\pm 0.003}$	$0.846_{\pm 0.004}$	$0.859_{\pm 0.004}$	-	$0.984_{\pm 0.003}$	$\mathbf{0.992}_{\pm 0.003}$
adult	$0.617_{\pm 0.001}$	$0.974_{\pm 0.003}$	$0.696_{\pm 0.003}$	$0.847_{\pm 0.003}$	$0.945_{\pm 0.003}$	$\underline{0.991}_{\pm 0.002}$	$\mathbf{0.997}_{\pm 0.001}$
bank	$0.810_{\pm 0.002}$	$\mathbf{0.995}_{\pm 0.001}$	$0.708_{\pm 0.002}$	$0.892_{\pm 0.004}$	$0.959_{\pm 0.004}$	$\underline{0.993}_{\pm 0.003}$	$0.984_{\pm 0.004}$
beijing	$0.916_{\pm 0.002}$	$0.957_{\pm 0.003}$	$0.928_{\pm 0.003}$	$\underline{0.980}_{\pm 0.001}$	$0.953_{\pm 0.004}$	$\mathbf{0.988}_{\pm 0.002}$	$\mathbf{0.988}_{\pm 0.002}$
churn	$0.957_{\pm 0.007}$	$\mathbf{0.975}_{\pm 0.005}$	$0.715_{\pm 0.007}$	$0.895_{\pm 0.012}$	$0.938_{\pm 0.005}$	$0.951_{\pm 0.010}$	$\underline{0.958}_{\pm 0.017}$
covertype	-	$0.940_{\pm 0.002}$	$0.753_{\pm 0.002}$	$0.787_{\pm 0.004}$	$\underline{0.957}_{\pm 0.004}$	$\mathbf{0.971}_{\pm 0.004}$	$0.918_{\pm 0.002}$

Appendix

Table 22: Big Benchmark:

\alpha

-Precision metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$\mathbf{0.460}_{\pm 0.003}$	$0.233_{\pm 0.002}$	$0.323_{\pm 0.003}$	-	$\underline{0.449}_{\pm 0.002}$	$0.436_{\pm 0.002}$
adult	$0.300_{\pm 0.001}$	$0.297_{\pm 0.003}$	$0.246_{\pm 0.002}$	$0.386_{\pm 0.003}$	$\mathbf{0.509}_{\pm 0.003}$	$0.467_{\pm 0.003}$	$\underline{0.478}_{\pm 0.002}$
bank	$\mathbf{0.461}_{\pm 0.002}$	$0.354_{\pm 0.003}$	$0.217_{\pm 0.002}$	$0.169_{\pm 0.002}$	$\underline{0.440}_{\pm 0.004}$	$0.417_{\pm 0.003}$	$0.393_{\pm 0.003}$
beijing	$\mathbf{0.613}_{\pm 0.002}$	$0.101_{\pm 0.002}$	$0.053_{\pm 0.001}$	$0.051_{\pm 0.001}$	$0.098_{\pm 0.001}$	$\underline{0.134}_{\pm 0.002}$	$0.113_{\pm 0.002}$
churn	$\mathbf{0.570}_{\pm 0.008}$	$0.146_{\pm 0.010}$	$0.003_{\pm 0.001}$	$0.062_{\pm 0.004}$	$0.094_{\pm 0.008}$	$0.179_{\pm 0.010}$	$\underline{0.257}_{\pm 0.012}$
covertype	-	$0.105_{\pm 0.001}$	$0.002_{\pm 0.000}$	$0.042_{\pm 0.001}$	$\mathbf{0.156}_{\pm 0.004}$	$\underline{0.146}_{\pm 0.003}$	$0.109_{\pm 0.002}$

Appendix

Table 23: Big Benchmark:

\beta

-Recall metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	Real	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	$0.390_{\pm 0.007}$	-	$\mathbf{0.376}_{\pm 0.008}$	$0.212_{\pm 0.005}$	$0.285_{\pm 0.009}$	-	$\underline{0.362}_{\pm 0.007}$	$0.334_{\pm 0.007}$
beijing	$0.803_{\pm 0.000}$	$\mathbf{0.671}_{\pm 0.005}$	$0.470_{\pm 0.010}$	$0.041_{\pm 0.019}$	$0.084_{\pm 0.034}$	$0.447_{\pm 0.015}$	$\underline{0.572}_{\pm 0.009}$	$0.407_{\pm 0.011}$
lending	$0.999_{\pm 0.000}$	$\mathbf{0.997}_{\pm 0.000}$	$0.885_{\pm 0.010}$	$-0.094_{\pm 0.061}$	$0.940_{\pm 0.006}$	$-1.623_{\pm 1.042}$	$0.984_{\pm 0.002}$	$\underline{0.992}_{\pm 0.001}$
news	$0.022_{\pm 0.004}$	$0.009_{\pm 0.025}$	$\mathbf{0.012}_{\pm 0.009}$	$0.006_{\pm 0.003}$	$-0.035_{\pm 0.001}$	$-0.573_{\pm 0.077}$	$-0.000_{\pm 0.020}$	$\underline{0.010}_{\pm 0.016}$
nmes	$0.177_{\pm 0.000}$	$0.103_{\pm 0.035}$	$0.143_{\pm 0.016}$	$-0.071_{\pm 0.023}$	$0.069_{\pm 0.010}$	$0.109_{\pm 0.033}$	$\mathbf{0.149}_{\pm 0.015}$	$\underline{0.148}_{\pm 0.019}$

Appendix

Table 24: Big Benchmark: MLE (

R^{2}

) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
adult	$0.929_{\pm 0.000}$	$0.912_{\pm 0.001}$	$0.908_{\pm 0.001}$	$0.891_{\pm 0.002}$	$0.880_{\pm 0.002}$	$0.911_{\pm 0.001}$	$\underline{0.913}_{\pm 0.001}$	$\mathbf{0.914}_{\pm 0.002}$
bank	$0.951_{\pm 0.002}$	$\underline{0.945}_{\pm 0.001}$	$0.939_{\pm 0.001}$	$0.878_{\pm 0.007}$	$0.931_{\pm 0.003}$	$0.941_{\pm 0.002}$	$0.938_{\pm 0.002}$	$\mathbf{0.947}_{\pm 0.001}$
churn	$0.988_{\pm 0.001}$	$\mathbf{0.980}_{\pm 0.002}$	$0.949_{\pm 0.012}$	$0.843_{\pm 0.019}$	$0.868_{\pm 0.014}$	$0.877_{\pm 0.025}$	$0.948_{\pm 0.015}$	$\underline{0.950}_{\pm 0.011}$
covertype	$0.964_{\pm 0.001}$	-	$\mathbf{0.898}_{\pm 0.002}$	$0.597_{\pm 0.010}$	$0.824_{\pm 0.002}$	$\underline{0.874}_{\pm 0.001}$	$0.852_{\pm 0.001}$	$0.834_{\pm 0.003}$
default	$0.774_{\pm 0.002}$	$\underline{0.760}_{\pm 0.003}$	$0.758_{\pm 0.003}$	$0.701_{\pm 0.005}$	$0.723_{\pm 0.004}$	$0.759_{\pm 0.003}$	$\mathbf{0.761}_{\pm 0.002}$	$0.757_{\pm 0.003}$
diabetes	$0.699_{\pm 0.002}$	$\mathbf{0.688}_{\pm 0.002}$	$\underline{0.678}_{\pm 0.003}$	$0.573_{\pm 0.003}$	$0.659_{\pm 0.001}$	-	$0.646_{\pm 0.004}$	$0.675_{\pm 0.002}$

Appendix

Table 25: Big Benchmark: MLE (ROCAUC) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
adult	$0.716_{\pm 0.002}$	$\mathbf{0.687}_{\pm 0.003}$	$0.649_{\pm 0.008}$	$0.646_{\pm 0.008}$	$0.620_{\pm 0.006}$	$0.678_{\pm 0.005}$	$0.675_{\pm 0.004}$	$\underline{0.682}_{\pm 0.004}$
bank	$0.605_{\pm 0.001}$	$\mathbf{0.574}_{\pm 0.010}$	$0.456_{\pm 0.016}$	$0.515_{\pm 0.012}$	$0.540_{\pm 0.014}$	$0.534_{\pm 0.010}$	$0.542_{\pm 0.012}$	$\underline{0.565}_{\pm 0.008}$
churn	$0.877_{\pm 0.001}$	$\mathbf{0.840}_{\pm 0.021}$	$0.698_{\pm 0.045}$	$0.535_{\pm 0.037}$	$0.582_{\pm 0.042}$	$0.340_{\pm 0.094}$	$0.705_{\pm 0.031}$	$\underline{0.713}_{\pm 0.042}$
covertype	$0.898_{\pm 0.003}$	-	$\mathbf{0.814}_{\pm 0.002}$	$0.347_{\pm 0.024}$	$0.737_{\pm 0.003}$	$\underline{0.789}_{\pm 0.002}$	$0.766_{\pm 0.003}$	$0.764_{\pm 0.003}$
default	$0.448_{\pm 0.004}$	$\mathbf{0.451}_{\pm 0.007}$	$0.368_{\pm 0.014}$	$0.247_{\pm 0.019}$	$0.338_{\pm 0.009}$	$\underline{0.439}_{\pm 0.008}$	$0.416_{\pm 0.011}$	$0.413_{\pm 0.013}$
diabetes	$0.588_{\pm 0.003}$	$\underline{0.574}_{\pm 0.004}$	$0.508_{\pm 0.007}$	$0.085_{\pm 0.011}$	$\mathbf{0.586}_{\pm 0.004}$	-	$0.435_{\pm 0.009}$	$0.560_{\pm 0.004}$

Appendix

Table 26: Big Benchmark: MLE (F1) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	Real	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	$0.774_{\pm 0.014}$	-	$\mathbf{0.783}_{\pm 0.016}$	$0.880_{\pm 0.017}$	$0.838_{\pm 0.017}$	-	$\underline{0.792}_{\pm 0.015}$	$0.809_{\pm 0.015}$
beijing	$0.441_{\pm 0.000}$	$\mathbf{0.571}_{\pm 0.005}$	$0.724_{\pm 0.007}$	$0.974_{\pm 0.010}$	$0.952_{\pm 0.017}$	$0.740_{\pm 0.010}$	$\underline{0.651}_{\pm 0.007}$	$0.766_{\pm 0.007}$
lending	$0.030_{\pm 0.000}$	$\mathbf{0.056}_{\pm 0.003}$	$0.338_{\pm 0.015}$	$1.043_{\pm 0.029}$	$0.245_{\pm 0.013}$	$1.588_{\pm 0.311}$	$0.126_{\pm 0.006}$	$\underline{0.090}_{\pm 0.006}$
news	$0.998_{\pm 0.002}$	$\underline{1.004}_{\pm 0.013}$	$\mathbf{1.003}_{\pm 0.005}$	$1.006_{\pm 0.002}$	$1.026_{\pm 0.000}$	$1.265_{\pm 0.031}$	$1.009_{\pm 0.010}$	$\underline{1.004}_{\pm 0.008}$
nmes	$0.985_{\pm 0.000}$	$1.028_{\pm 0.020}$	$1.005_{\pm 0.010}$	$1.123_{\pm 0.012}$	$1.048_{\pm 0.006}$	$1.024_{\pm 0.019}$	$\mathbf{1.001}_{\pm 0.009}$	$\underline{1.002}_{\pm 0.011}$

Appendix

Table 27: Big Benchmark: MLE (RMSE) metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined . Real indicates MLE results when training on real data.

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$0.980_{\pm 0.006}$	$\underline{0.998}_{\pm 0.002}$	$0.994_{\pm 0.006}$	-	$\mathbf{0.999}_{\pm 0.002}$	$\underline{0.998}_{\pm 0.002}$
adult	$0.948_{\pm 0.006}$	$0.914_{\pm 0.007}$	$\underline{0.994}_{\pm 0.005}$	$\mathbf{0.996}_{\pm 0.004}$	$0.970_{\pm 0.007}$	$0.980_{\pm 0.008}$	$0.984_{\pm 0.007}$
bank	$0.882_{\pm 0.006}$	$0.915_{\pm 0.007}$	$0.982_{\pm 0.007}$	$0.984_{\pm 0.010}$	$0.978_{\pm 0.008}$	$\mathbf{0.993}_{\pm 0.006}$	$\underline{0.989}_{\pm 0.006}$
beijing	$0.734_{\pm 0.005}$	$0.940_{\pm 0.008}$	$\underline{0.981}_{\pm 0.007}$	$0.978_{\pm 0.011}$	$\mathbf{0.987}_{\pm 0.007}$	$0.974_{\pm 0.009}$	$0.975_{\pm 0.006}$
churn	$0.834_{\pm 0.032}$	$0.933_{\pm 0.031}$	$\underline{0.982}_{\pm 0.015}$	$0.965_{\pm 0.025}$	$\underline{0.982}_{\pm 0.018}$	$\mathbf{0.983}_{\pm 0.016}$	$0.943_{\pm 0.024}$
covertype	-	$0.969_{\pm 0.004}$	$0.993_{\pm 0.010}$	$0.993_{\pm 0.005}$	$\mathbf{0.997}_{\pm 0.003}$	$\underline{0.995}_{\pm 0.004}$	$0.994_{\pm 0.005}$

Appendix

Table 28: Big Benchmark: DCR metric scores for all datasets. Average and standard devation

(\mu\pm\sigma)

over 20 sampling seeds. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$89.414$	$16.844$	$\underline{8.263}$	-	$684.166$	$\mathbf{2.877}$
adult	-	$4.784$	$6.400$	$\underline{2.722}$	$25.883$	$70.789$	$\mathbf{0.077}$
bank	-	$10.545$	$6.309$	$\underline{2.969}$	$26.113$	$89.043$	$\mathbf{0.057}$
beijing	-	$5.931$	$5.177$	$\underline{2.917}$	$22.569$	$73.557$	$\mathbf{0.044}$
churn	-	$0.341$	$0.416$	$\underline{0.295}$	$1.874$	$45.802$	$\mathbf{0.022}$
covertype	-	$143.582$	$26.471$	$\underline{12.510}$	$40.873$	$950.345$	$\mathbf{0.948}$

Appendix

Table 29: Big Benchmark: training time in minutes for all datasets. Best results per dataset are in bold , second best underlined .

model	SMOTE	ARF	CTGAN	TVAE	TabDDPM	TabSyn	XGenB-AR
acsincome	-	$101.964_{\pm 3.579}$	$\underline{2.003}_{\pm 0.038}$	$\mathbf{1.454}_{\pm 0.053}$	-	$37.829_{\pm 0.245}$	$8.898_{\pm 0.276}$
adult	$12.889_{\pm 0.179}$	$38.817_{\pm 0.605}$	$\underline{0.449}_{\pm 0.006}$	$\mathbf{0.264}_{\pm 0.007}$	$141.720_{\pm 1.729}$	$14.286_{\pm 0.091}$	$2.481_{\pm 0.073}$
bank	$12.391_{\pm 0.339}$	$39.714_{\pm 0.669}$	$\underline{0.453}_{\pm 0.005}$	$\mathbf{0.261}_{\pm 0.005}$	$136.956_{\pm 0.468}$	$11.998_{\pm 0.085}$	$1.645_{\pm 0.048}$
beijing	$6.272_{\pm 0.050}$	$4.916_{\pm 0.154}$	$\underline{0.294}_{\pm 0.007}$	$\mathbf{0.166}_{\pm 0.004}$	$70.876_{\pm 0.665}$	$9.149_{\pm 0.012}$	$1.020_{\pm 0.008}$
churn	$0.203_{\pm 0.003}$	$1.520_{\pm 0.047}$	$\underline{0.059}_{\pm 0.001}$	$\mathbf{0.051}_{\pm 0.001}$	$10.963_{\pm 0.074}$	$0.611_{\pm 0.003}$	$0.218_{\pm 0.004}$
covertype	-	$519.985_{\pm 9.916}$	$\underline{3.092}_{\pm 0.049}$	$\mathbf{1.689}_{\pm 0.035}$	$1295.988_{\pm 182.294}$	$39.751_{\pm 0.205}$	$9.042_{\pm 0.350}$

Appendix

Table 30: Big Benchmark: inference time in seconds for all datasets. Best results per dataset are in bold , second best underlined .

Synthetic Tabular Data Generation

Synthetic tabular data generation creates artificial records that preserve useful feature relationships while limiting exposure of sensitive data.

See all 98 papers

Mar 11, 2026

ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning

Deep generative models can help with data scarcity and privacy by producing synthetic training data, but they struggle in low-data, imbalanced tabular settings to fully learn the complex data distribution. We argue that striving for the full joint distribution could be overkill; for greater data efficiency, models should prioritize learning the conditional distribution

P(y\mid \bm{X})

, as suggested by recent theoretical analysis. Therefore, we overcome this limitation with ReTabSyn, a Reinforced Tabular Synthesis pipeline that provides direct feedback on feature correlation preservation during synthesizer training. This objective encourages the generator to prioritize the most useful predictive signals when training data is limited, thereby strengthening downstream model utility. We empirically fine-tune a language model-based generator using this approach, and across benchmarks with small sample sizes, class imbalance, and distribution shift, ReTabSyn consistently outperforms state-of-the-art baselines. Moreover, our approach can be readily extended to control various aspects of synthetic tabular data, such as applying expert-specified constraints on generated observations.

Xiaofeng Lin, Seungbae Kim, Zhuoya Li +3

University of California, Los Angeles, USA · University of South Florida, USA · Cisco, USA +1

Synthetic Tabular Data Generation Tabular Data Synthesis Conditional Generative Modeling RL for Generative Models

Ask PDF

Sep 30, 2026

CDMD: A Cross-Dataset Mixed-Type Diffusion Model for Tabular Data

Generative models for tabular data are typically trained separately for each dataset, limiting knowledge transfer and requiring the storage of many specialized models. In this paper, we introduce CDMD, a tabular diffusion model trained jointly across heterogeneous datasets with different schemas and variable numbers of numerical and categorical features. Unlike existing cross-dataset tabular diffusion models that operate in continuous representation spaces, CDMD defines diffusion directly over the mixed-type feature space and is trained end-to-end. To accommodate heterogeneous categorical domains, we introduce a schema-restricted reverse-process parameterization for masked diffusion models, in which the output space dynamically adapts to each feature's vocabulary. We then compose numerical and categorical feature-level diffusion processes into a schema-dependent row-level process. A shared schema-aware Transformer denoiser captures dependencies between features and parameterizes the reverse process across varying schemas. On seven real-world datasets, a single jointly trained CDMD achieves the highest average generation quality among strong single-dataset and cross-dataset baselines, while using substantially fewer total parameters than the collection of separately trained models. Furthermore, pre-training on a corpus of 337 datasets improves generation on previously unseen datasets under both limited target data and limited adaptation epochs. These results demonstrate the potential of direct mixed-type diffusion for shared and transferable tabular data generation. Our code is available at https://github.com/ketatam/cdmd.

Mohamed Amine Ketata, Maximilian Schambach, Stephan Günnemann

Munich Data Science Institute, Technical University of Munich, Germany · SAP SE, Germany

Masked Diffusion Models Tabular Diffusion Models Synthetic Tabular Data Generation Diffusion Models

Ask PDF

Mar 31, 2026

SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection

Deep generative models for tabular data (GANs, diffusion models, and LLM-based generators) exhibit highly non-uniform behavior across datasets; the best-performing synthesizer family depends strongly on distributional stressors such as long-tailed marginals, high-cardinality categorical, Zipfian imbalance, and small-sample regimes. This brittleness makes practical deployment challenging, especially when users must balance competing objectives of fidelity, privacy, and utility. We study {intent-conditioned tabular synthesis selection}: given a dataset and a user intent expressed as a preference over evaluation metrics, the goal is to select a synthesizer that minimizes regret relative to an intent-specific oracle. We propose {stress profiling}, a synthesis-specific meta-feature representation that quantifies dataset difficulty along four interpretable stress dimensions, and integrate it into {SYNTHONY}, a selection framework that matches stress profiles against a calibrated capability registry of synthesizer families. Across a benchmark of 7 datasets, 10 synthesizers, and 3 intents, we demonstrate that stress-based meta-features are highly predictive of synthesizer performance: a

k

NN selector using these features achieves strong Top-1 selection accuracy, substantially outperforming zero-shot LLM selectors and random baselines. We analyze the gap between meta-feature-based and capability-based selection, identifying the hand-crafted capability registry as the primary bottleneck and motivating learned capability representations as a direction for future work.

Hochan Son, Xiaofeng Lin, Jason Ni +1

Department of Statistics, University of California, Los Angeles · Department of Mathematics, University of California, Los Angeles

Synthetic Data Selection Tabular Data Synthesis Generative Modeling

Ask PDF

Abstract

Synthetic Tabular Data Generation

Explore similar work

ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning

CDMD: A Cross-Dataset Mixed-Type Diffusion Model for Tabular Data

SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection