P-TA: Using Proximal Policy Optimization to Enhance Tabular Data Augmentation Via Large Language Models

A multitude of industries depend on accurate and reasonable tabular dataaugmentation for their business processes. Contemporary methodologies ingenerating tabular data revolve around utilizing Generative AdversarialNetworks (GAN) or fine-tuning Large Language Models (LLM). However, GAN-basedapproaches are documented to produce samples with common-sense errorsattributed to the absence of external knowledge. On the other hand, LLM-basedmethods exhibit a limited capacity to capture the disparities betweensynthesized and actual data distribution due to the absence of feedback from adiscriminator during training. Furthermore, the decoding of LLM-basedgeneration introduces gradient breakpoints, impeding the backpropagation ofloss from a discriminator, thereby complicating the integration of these twoapproaches. To solve this challenge, we propose using proximal policyoptimization (PPO) to apply GANs, guiding LLMs to enhance the probabilitydistribution of tabular features. This approach enables the utilization of LLMsas generators for GANs in synthesizing tabular data. Our experimentsdemonstrate that PPO leads to an approximately 4% improvement in the accuracyof models trained on synthetically generated data over state-of-the-art acrossthree real-world datasets.