TabularARGN: A Flexible and Efficient Auto-Regressive Framework for Generating High-Fidelity Synthetic Data

TL;DR

TabularARGN tackles the challenge of generating high-fidelity synthetic data for mixed-type tabular datasets while balancing privacy and efficiency. It adopts a shallow auto-regressive architecture that models the joint distribution via $p(oldsymbol{x}) = \prod_{i=1}^D p(x_i \mid x_{<i})$, extended to any-order conditioning across columns and, for sequences, across time steps with an LSTM-based history encoder. Key contributions include data-safe encoding that maps all features to categorical sub-columns, a flat-table model with any-order conditioning and shared embeddings, and a sequential-table model that can handle variable-length, irregularly spaced sequences and two-table conditioning with flat context; the framework also supports differential privacy via DP-SGD. Empirically, TabularARGN achieves state-of-the-art-like synthetic data quality with substantially faster training and inference than competitive baselines, across flat and sequential benchmarks, while maintaining robustness under DP and demonstrating practical scalability. The open-source implementation enables industry-scale deployment and flexible generation tasks such as imputation and fairness-aware sampling, signaling a significant advance for private yet utility-preserving synthetic data generation in real-world settings.

Abstract

Synthetic data generation for tabular datasets must balance fidelity, efficiency, and versatility to meet the demands of real-world applications. We introduce the Tabular Auto-Regressive Generative Network (TabularARGN), a flexible framework designed to handle mixed-type, multivariate, and sequential datasets. By training on all possible conditional probabilities, TabularARGN supports advanced features such as fairness-aware generation, imputation, and conditional generation on any subset of columns. The framework achieves state-of-the-art synthetic data quality while significantly reducing training and inference times, making it ideal for large-scale datasets with diverse structures. Evaluated across established benchmarks, including realistic datasets with complex relationships, TabularARGN demonstrates its capability to synthesize high-quality data efficiently. By unifying flexibility and performance, this framework paves the way for practical synthetic data generation across industries.

Figures (5)

• Figure 1: (a) Model components and information flow in the training phase of a three-column TabularARGN flat model with the current column order [1,2,3]. Input features $x_{i}$ are embedded and sent through the permutation masking layer to condition predictions on preceding columns. The permutation masking layer randomly shuffles the column order for each training batch. (b) Model components and information flow in the generation phase. The input to the model and starting point of the generation is a vector of zeros (blue) triggering the successive generation of synthetic features. Due to the permutation of column orders during training, any column order can be realized in the generation phase.
• Figure 2: Model components and information flow in the training phase of a three-column TabularARGN sequential model with a two-column flat context. The current column order is [1,2,3]. Input features $x_{i;1:T}$ are embedded and sent through both the permutation masking layer and the history encoder block (blue) to condition predictions on preceding columns and time steps, respectively. For conditional sequence generation, flat context features $x_{ci}$ are ingested by the context processor (orange). They are encoded, compressed, and provided as context for the prediction of all columns of the sequential training data.
• Figure 3: Training time (top row) and Accuracy (bottom row) for the flat Adult (left column) and sequential Baseball (right column) data set. Reported values are averages over five full training and generation runs, error bars indicate the minimum and maximum values. As UnmaskingTrees is a non-Deep Learning approach, it is run on CPU. Training of TabularARGN on the Adult with a CPU takes approximately 2.5 minutes.
• Figure 4: Accuracy (top row) and Training Time (bottom row) for the flat datasets Adult, ACS-Income, Default, and Shoppers and various generative models - TabularARGN, TabularARGN with DP, TabMT, TabSyn, CTGAN, STaSy, and UnmaskingTree. Reported values are averages over five full training and generation runs, error bars indicate the minimum and maximum values. As UnmaskingTrees is a non-Deep Learning approach, it is run on CPU. For comparison, Training of TabularARGN on the Adult with a CPU takes approximately 2.5 minutes.
• Figure 5: Accuracy (top row) and Training Time (bottom row) for the sequential datasets Baseball, California, and Berka and various generative models - TabularARGN, TabularARGN with DP, ClavaDDPM, RCTGAN, RealTabFormer, and SDV/ParSynthesizer. Reported values are averages over at least three full training and generation runs, error bars indicate the minimum and maximum values. RCTGAN and SDV/ParSynthesizer fail to train on the Berka and California datasets, respectively, due to out-of-memory (OOM) issues.