Title: TabChange: Achieving Precise Attribute Modifications in Tabular Data

Abstract: Adjusting a single attribute within tabular data frequently results in unnatural samples, as such alterations can disrupt the inherent correlations between variables. To be viable, any modified instance must remain close to the original while still appearing realistic. This study tackles the difficulty of producing such modified instances by highlighting significant shortcomings in current methodologies. Specifically, generative models often lack support for instance-level attribute editing, while approaches like CVAE tend to preserve attribute details within the latent space, causing superfluous changes.

To overcome these issues, we introduce TabChange, a novel technique that examines the dependency between the target attribute and the remaining features in the dataset. The method employs a dual strategy: if the correlation is weak, it directly inverts the attribute; if the correlation is strong, it utilizes an adversarial framework to strip attribute-specific information from the latent representation. This process ensures precise edits, applying only the adjustments necessary to preserve the data's natural structure.

Evaluations across seven distinct datasets demonstrate that TabChange produces counterfactuals with naturalness levels comparable to existing methods, while remaining closer to the original instances. Consequently, this approach yields a greater quantity of valid counterfactuals and reduces the occurrence of invalid ones relative to baseline models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC