Quasi-Synthetic Data Generation Framework

• Quasi-synthetic data generation frameworks are methods that create synthetic records by anchoring generation to real evidence such as source corpora, aggregated statistics, or geometric structures.
• They employ diverse techniques including Gaussian copulas, Riemannian sampling, and modular DAG-based pipelines to maintain statistical consistency with the observed data.
• These frameworks integrate task-aware filtering and profile guidance to effectively balance privacy, fidelity, and utility, addressing key limitations of conventional synthetic data approaches.

Quasi-synthetic data generation frameworks produce synthetic records whose content is not created from unconstrained priors, but is instead anchored to real evidence such as source corpora, aggregated statistics, fitted data distributions, or domain-specific geometric structure. In recent work, this anchor appears in several forms: conversational pipelines that transform existing OASST-style or HuggingFace datasets into training-ready SFT or DPO records (Pradhan et al., 21 Aug 2025), macro-to-microdata systems that reconstruct individual-level populations from aggregated tables (Acharya et al., 2022, Li et al., 2020), model-based resampling of proprietary graphs and tabular datasets (Darabi et al., 2022, Ling et al., 2023), and Riemannian sampling schemes that expand a small seed set of signature descriptors into synthetic writer populations (Zois et al., 24 Sep 2025). Across these variants, the common structure is a real-data anchor, a generation mechanism, and a constraint, filtering, or evaluation layer that determines whether the outputs are realistic enough, safe enough, and useful enough for the intended downstream use.

1. Conceptual scope and recurring modes

The literature does not present one universal formal taxonomy of “quasi-synthetic” data, but several recurring modes are explicit. GraSP distinguishes a fully synthetic “data-less mode” from pipelines that begin with source data and then transform, augment, filter, relabel, and serialize them into new artifacts; in that setting, quasi-synthetic generation corresponds to source-conditioned graph execution rather than prompt-only synthesis (Pradhan et al., 21 Aug 2025). GenSyn and SYNC define a different mode: the target individual-level dataset is never observed directly, and synthetic microdata are reconstructed from lower-resolution marginals, cross-tabs, or other aggregate evidence (Acharya et al., 2022, Li et al., 2020). In graph learning, the outputs are newly sampled graphs, yet they remain statistically anchored to a proprietary source graph through fitted topology, feature, and alignment models, which the authors explicitly describe as quasi-synthetic in character (Darabi et al., 2022). A related tabular formulation appears in local or neighborhood-based synthesis, where synthetic rows are generated from real-data partitions or leaf clusters and therefore preserve local geometry without copying original rows verbatim (Ling et al., 2023).

Mode	Real anchor	Representative frameworks
Source-conditioned transformation	Existing conversational corpora or OASST trees	GraSP (Pradhan et al., 21 Aug 2025)
Aggregate-to-microdata reconstruction	Marginals, cross-tabs, auxiliary regional summaries	GenSyn (Acharya et al., 2022), SYNC (Li et al., 2020)
Model-based resampling	Proprietary graph or tabular dataset fit by a generator	Large-scale graph framework (Darabi et al., 2022), Howso/RRP (Ling et al., 2023)
Partitioned or profile-guided synthesis	Real columns or sample profiles split before generation	DGMs (Lautrup et al., 25 Jul 2025), data-centric profile-guided generation (Hansen et al., 2023)
Seed-set geometric expansion	Small genuine set on a structured manifold	SPD signature framework (Zois et al., 24 Sep 2025)

A recurring implication is that quasi-synthetic frameworks are defined less by one generator family than by how they preserve a statistical or semantic relationship to observed reality. In some cases, the real anchor is a distribution over rows; in others, it is only an aggregate constraint set, a conversation tree, or a manifold-valued seed population.

2. Data anchors and mathematical constructions

Aggregate-to-microdata frameworks are the clearest examples of quasi-synthetic reconstruction. GenSyn assumes target-region univariate marginals $D1$, target-region bi- or multivariate tables $D2$, and auxiliary-region univariate distributions $D3$. It combines a target-specific conditional model $p_1$, an auxiliary Gaussian-copula model $p_2$, and a maximum-entropy correction stage that enforces the target marginals. The final optimization uses the combined prior $p$ and solves a minimum cross-entropy problem over tuple weights $w_i$, with low-probability tuples pruned by a threshold $\tau$; in the experiments, $\tau = 1/N$ gave the best trade-off between tractability and rare-profile retention (Acharya et al., 2022).

SYNC formalizes the same family of problems from the perspective of downscaling. For each aggregation unit $m$ and feature $D2$0, the observed input is an aggregate

$D2$1

while the desired output is the unobserved individual-level set $D2$2. SynC fits a Gaussian copula to each batch of aggregate variables, samples high-resolution subsets by latent Gaussian transformation, merges batches through predictive models conditioned on a predefined core feature set, and finally rescales the generated records so that categorical class counts and continuous means agree with the original coarse summaries (Li et al., 2020).

A different mathematical anchor appears in writer-independent offline signature verification, where the data live on the SPD manifold

$D2$3

The quasi-synthetic mechanism is a Riemannian Gaussian mixture

$D2$4

where the Riemannian centers $D2$5 are interpreted as synthetic writers and $D2$6 as their variability. Synthetic genuine samples are drawn from $D2$7, while synthetic forgeries are drawn from $D2$8 with $D2$9 for “Hard” training and $D3$0 for “Soft” training (Zois et al., 24 Sep 2025). This construction is quasi-synthetic because a small genuine seed set initializes the mixture, after which training proceeds entirely on synthetic SPD pairs.

These examples illustrate a general pattern: quasi-synthetic generation is often governed by explicit constraints or geometry rather than by an unconstrained end-to-end likelihood. Marginal preservation, manifold structure, and aggregate consistency are not post hoc diagnostics alone; they are part of the generative definition.

3. System architectures and compositional pipelines

Many frameworks define quasi-synthetic generation as a systems problem rather than as a single probabilistic model. GraSP is exemplary in this respect. It uses a modular, configuration-driven pipeline with four layers—Data I/O, Graph Construction, Execution Engine, and Structured Output and Resumability—in which workflows are declared in YAML as DAG-like graphs with conditional logic, reusable subgraphs, and nodes for LLM calls, Python logic, judges, or tool-using agents. The framework supports both source-conditioned and data-less execution, dual-stage quality tagging with heuristic rules plus LLM-based evaluation, OASST-compatible output mapping, and resumable generation across backends such as OpenAI, Mistral, vLLM, TGI, OLLAMA, and Azure (Pradhan et al., 21 Aug 2025).

Another compositional strategy is to mix generators rather than commit to one. SC-GOAT defines a pool $D3$1, tunes each synthesizer with downstream validation loss, and then learns mixture weights $D3$2 so that the final synthetic dataset concatenates $D3$3 rows from each generator. The framework is explicitly supervised: the outer optimization chooses $D3$4 and $D3$5 to optimize downstream AUC, and the learned solution may be a true mixture or a corner solution that collapses to a single generator when that is empirically optimal (Nakamura-Sakai et al., 2023).

Disjoint Generative Models push compositionality further by partitioning the feature set column-wise into disjoint subsets, training separate generators on each subset, and rejoining the independently generated partitions without shared identifiers. The main joining strategy is validator-based: a classifier is trained to distinguish authentic full rows from randomly joined partitions, then candidate synthetic joins are accepted when the validator score exceeds a threshold $D3$6. This yields a modular privacy–utility control knob through partition structure, model choice per partition, and join threshold (Lautrup et al., 25 Jul 2025).

Cross-table foundation models represent another architectural generalization. CTSyn combines an aggregator that maps heterogeneous rows from many tables into a unified latent vector, a conditional latent diffusion model for sampling that latent space, and type-specific decoders that reconstruct numerical and categorical values from $D3$7, where $D3$8 is metadata embedding, $D3$9 is column-name embedding, and $p_1$0 is the row latent. Because the aggregator and decoders are reused across tables, CTSyn turns cross-table pretraining into a reusable synthetic-data backbone rather than a per-table synthesizer (Lin et al., 2024).

In large-scale graph generation, modularity appears as a decomposition into a structural generator, a feature generator, and an aligner. The structure module uses a generalized Kronecker/R-MAT-style generator, the feature module treats node and edge attributes as tabular samples, and the aligner restores structure–feature dependence through predictors such as XGBoost. The resulting graphs are synthetic in form but data-grounded in topology, feature distributions, and structure–attribute couplings learned from the proprietary source graph (Darabi et al., 2022).

4. Task-aware filtering, profile guidance, and synthetic boosting

A major development in quasi-synthetic design is the shift from unconditional generation toward task-aware selection. One route is profile-guided generation. In a data-centric benchmark, real training data are first scored by Cleanlab, Data-IQ, or Data Maps and partitioned into profiles such as easy, ambiguous, and hard. Separate generators are then trained on each profile subset, synthetic data are recombined while preserving profile proportions, and a postprocessing option removes synthetic hard examples. In that benchmark, profile-based preprocessing and/or postprocessing improved classification in 13 of 15 evaluations, feature selection in 12 of 15, and model selection by as much as 92% in one setting, even though fidelity scores could decrease slightly (Hansen et al., 2023).

A second route is statistical filtering after large-model generation. In the framework for boosting statistic learning with pretrained large models, Stable Diffusion XL Refiner generates large candidate pools, but only a subset is retained. For tabular data, generated candidates are filtered through transferability detection using glmtrans and hdtrd, and the selected source fraction is chosen by

$p_1$1

For images, generated samples are ranked in latent space using Wasserstein distance, with retained fractions such as the top 80% on MNIST and top 60% on CIFAR-10 and ISIC. The core result is that only a limited subset of generated samples is actually boostable for downstream prediction (Jiang et al., 8 May 2025).

A third route is to treat synthetic size itself as a tunable quantity. Synthetic Data Generation for Analytics formalizes the “generational effect”: downstream error on synthetic data can decrease as synthetic volume grows, but may later stabilize or rise because generation error accumulates. The optimal synthetic size is the reflection point

$p_1$2

This paper does not define quasi-synthetic generation via a new model family; instead, it defines a risk-aware workflow in which pre-trained or fine-tuned generators are used for volume expansion only up to the point where added synthetic mass still improves prediction or inference (Shen et al., 2023).

Taken together, these frameworks replace the older assumption that synthetic data quality is a monotone function of sample count or marginal fidelity. The emerging design principle is selective integration: profile, rank, filter, or size-control the synthetic pool before it enters downstream analytics.

5. Evaluation, privacy, and utility governance

Evaluation has become a first-class component of quasi-synthetic frameworks. The most explicit formulation is the Representativeness–Privacy–Utility framework, which treats synthetic trip data as a $p_1$3 matrix of dimensions and levels. Representativeness $p_1$4 is measured at record, group, and population levels; privacy is split into record-level disclosure, group-level exposure, and population-level leakage; utility is defined relative to downstream tasks. The framework uses KL divergence, Jensen–Shannon divergence, and Wasserstein distance for representativeness, a customized Random Forest membership inference attack for record-level privacy, k-nearest-neighbor distance ratios for group- and population-level privacy, and clustering plus TSTR prediction for utility. In the Hong Kong Octopus benchmark, the main conclusions are that synthetic data do not inherently guarantee privacy, there is no “one-size-fits-all” model, and CTGAN provided the most balanced trade-off in that particular setup (Wu et al., 28 Oct 2025).

For unlabeled tabular settings, another evaluation framework defines a five-stage procedure: generate synthetic candidates, screen them with Diagnostic Validity, score them using Wasserstein–Cramér’s V, Novelty, domain classifier, and anomaly detection, and then rank generators with Friedman Aligned-Ranks followed by Finner post-hoc testing. The motivation is that heterogeneous metrics often disagree, and direct averaging of scores lacks theoretical support. This framework therefore treats model selection itself as a nonparametric statistical problem rather than as a leaderboard of unrelated metrics (Livieris et al., 2024).

A related line of work studies privacy–utility–scalability trade-offs directly. In a comparison of Howso, a recursive random projection framework, DataSynthesizer, and Synthetic Data Vault variants, DataSynthesizer achieved the strongest privacy scores, the proposed random projection framework had the highest utility and fastest scalability, and Howso had the best overall balance. The proposed recursive random projection method is especially close to quasi-synthetic generation because it first partitions real data into small leaf clusters and then generates new points by mutation and crossover within those local neighborhoods (Ling et al., 2023).

Two misconceptions are repeatedly rejected by these evaluation papers. First, high statistical similarity does not imply high downstream utility: in one data-centric benchmark, a Bayesian network achieved the highest inverse KL-divergence while performing poorly on classification, model selection, and feature selection (Hansen et al., 2023). Second, synthetic data are not automatically private: high similarity can be evidence of overfitting or memorization rather than of safety (Wu et al., 28 Oct 2025).

6. Limitations, trade-offs, and open directions

The literature converges on several recurring limitations. One is the persistence of privacy–utility trade-offs. Privacy-enhanced models may reduce leakage but often degrade fidelity and task utility, while high-fidelity models can remain vulnerable to membership or linkage-style attacks. Disjoint Generative Models make this tension explicit: more partitions generally improve privacy but degrade utility, and validator-based rejoining restores some dependence at the cost of losing clean end-to-end differential privacy guarantees (Lautrup et al., 25 Jul 2025).

A second limitation is domain specificity in the constraint layer. Public-transport trip data require logic checks such as valid origin–destination pairs and time ranges (Wu et al., 28 Oct 2025). GraSP depends on OASST-style conversational structures, schema validation, and text-only outputs even though multimodal input is supported (Pradhan et al., 21 Aug 2025). The signature framework presupposes SPD covariance descriptors and Riemannian Gaussian sampling (Zois et al., 24 Sep 2025). GenSyn is mainly categorical and depends on harmonized cross-tabs and auxiliary regions, while its tuple space grows exponentially and must be pruned heuristically (Acharya et al., 2022). These examples suggest that no universal constraint language yet spans the full quasi-synthetic landscape.

A third limitation is partial coverage of data types and global semantics. CTSyn explicitly implements numerical and categorical decoding, but the paper does not provide equally concrete mechanisms for all other tabular types or formal privacy guarantees (Lin et al., 2024). The large-scale graph framework decouples structure and features and then repairs dependence by alignment; the authors note that this can be inappropriate for domains such as molecular graphs where topology and attributes are tightly coupled (Darabi et al., 2022). GraSP warns about model collapse risks from recursively training on generated data and notes limited cross-sample reasoning for global diversity control (Pradhan et al., 21 Aug 2025).

A plausible synthesis of these directions is a layered quasi-synthetic framework in which cross-source or cross-table generators produce candidates, profile- or transfer-based filters select useful subsets, and multi-level evaluation audits representativeness, privacy, and utility before release. The surveyed papers do not present that single unified system, but collectively they make its design logic explicit: real anchors are indispensable, generation is only one stage in the pipeline, and deployment requires deliberate control of constraints, subgroup behavior, and task-specific validity rather than faith in syntheticity alone.