ABSTRACT
Recently, generative models have been gradually emerging into the extended dataset field, showcasing their advantages. However, when it comes to generating tabular data, these models often fail to satisfy the constraints of numerical columns, which cannot generate high-quality datasets that accurately represent real-world data and are suitable for the intended downstream applications. Responding to the challenge, we propose a tabular data generation framework guided by downstream task optimization (TDGGD). It incorporates three indicators into each time step of diffusion generation, using gradient optimization to align the generated fake data. Unlike the traditional strategy of separating the downstream task model from the upstream data synthesis model, TDGGD ensures that the generated data has highly focused columns feasibility in upstream real tabular data. For downstream task, TDGGD strikes the utility of tabular data over solely pursuing statistical fidelity. Through extensive experiments conducted on real-world tables with explicit column constraints and tables without explicit column constraints, we have demonstrated that TDGGD ensures increasing data volume while enhancing prediction accuracy. To the best of our knowledge, this is the first instance of deploying downstream information into a diffusion model framework.
