International Journal of Science, Engineering and TechnologyAuthors: Shiv Hari Tewari
Abstract: The creation of high-quality synthetic data has become essential in modern machine learning applications, tackling several important challenges in data science. In many real-world situations, organizations experience significant constraints when using actual datasets. These limitations can include having insufficient amounts of data, containing sensitive personal information, or displaying biases that distort analysis results. Synthetic data generation provides an effective solution by producing artificial datasets that statistically resemble real-world data while avoiding these issues. This research specifically aims to improve synthetic data generation through the fine-tuning of Generative Pre-trained Transformer (GPT) models. These models are part of deep learning architectures known for their ability to understand and generate complex patterns. This framework consists of: -Statistical similarity metrics, like Jensen-Shannon divergence, which measure how closely the synthetic data distribution matches the original one. – Propensity score analysis, a technique to check if classifiers can tell apart real and synthetic samples, with indistinguishable samples indicating higher quality. -Machine learning efficacy tests that check the practical usefulness of synthetic data by training predictive models and comparing their performance to models trained on original data. Our findings show that fine-tuned GPT models perform better in all evaluation areas. They not only replicate the complex statistical properties of real-world datasets more accurately but also allow predictive models trained on synthetic data to perform at levels similar to those trained on real datasets. This result is significant, as it suggests that well-tuned synthetic data can act as an effective substitute for real data in many analytical settings. This is especially relevant in sensitive areas where preserving privacy is critical, such as healthcare or finance, or in cases where model generalization is needed despite limited training data. Our approach presents a practical solution by showing that synthetic data can maintain both statistical integrity and functional usefulness. This research opens new avenues for data sharing, collaborative studies, and the creation of more inclusive and representative machine learning systems, addressing important ethical and practical challenges tied to using real-world data.
