Creating Artificial Students that Never Existed: Leveraging Large Language Models and CTGANs for Synthetic Data Generation
This addresses data scarcity for learning analytics researchers, but it is incremental as it applies existing generative methods to a new domain.
The study tackled the problem of limited access to quality student data for learning analytics due to privacy concerns by generating synthetic tabular data using CTGAN and LLMs like GPT2, DistilGPT2, and DialoGPT, demonstrating strong potential to produce high-quality datasets that resemble real data.
In this study, we explore the growing potential of AI and deep learning technologies, particularly Generative Adversarial Networks (GANs) and Large Language Models (LLMs), for generating synthetic tabular data. Access to quality students data is critical for advancing learning analytics, but privacy concerns and stricter data protection regulations worldwide limit their availability and usage. Synthetic data offers a promising alternative. We investigate whether synthetic data can be leveraged to create artificial students for serving learning analytics models. Using the popular GAN model CTGAN and three LLMs- GPT2, DistilGPT2, and DialoGPT, we generate synthetic tabular student data. Our results demonstrate the strong potential of these methods to produce high-quality synthetic datasets that resemble real students data. To validate our findings, we apply a comprehensive set of utility evaluation metrics to assess the statistical and predictive performance of the synthetic data and compare the different generator models used, specially the performance of LLMs. Our study aims to provide the learning analytics community with valuable insights into the use of synthetic data, laying the groundwork for expanding the field methodological toolbox with new innovative approaches for learning analytics data generation.