Generative models are increasingly used to produce synthetic tabular data for applications including fraud detection, privacy-preserving analytics, and stress-testing downstream models. A fundamental challenge is that statistical fidelity and semantic validity are not the same objective: a model can reproduce realistic-looking data that still violates domain constraints. A new paper from UvA researchers introduces a framework that addresses this gap without requiring explicit rule specification.
In critical domains such as finance, healthcare, and manufacturing, the quality of synthetic data is not merely a matter of statistical accuracy. Generative models can faithfully reproduce marginal distributions while still producing records that violate fundamental domain constraints, for example transaction amounts inconsistent with account types, or risk flags incompatible with counterparty classifications. Such errors represent implausible or legally incoherent scenarios with real-world consequences.
LSR (Latent Semantic Regularization) addresses this by learning implicit constraint structure directly from data, without requiring explicit rule specification, feasibility oracles, or curated negative examples. The framework operates in two phases: first, a validator is pre-trained as a variational autoencoder with a normalizing flow prior; then, frozen, it is incorporated as a semantic regularization term during training of the generative model. This steers the generation process toward semantically plausible outputs without requiring domain knowledge to be explicitly encoded.
For financial data synthesis, where constraint violations carry real consequences, LSR provides rigorous evaluation infrastructure and inductive bias suited to AI applications in audit and compliance. The paper is authored by Saba Amiri, Carlijn Nijhuis, Eric Nalisnick, Adam Belloum, Sander Klous, and Leon Gommans, and will be presented at the Symposium on Probabilistic Machine Learning in Seoul, South Korea, in July 2026.
Read also
