Minority class augmentation in tabular data using generative adversarial network models

Date

2023-9

Author

Polat, Esranur

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

180
views

121
downloads

Cite This

In the rapidly developing technology environment, the interaction between advancing technology and the exponential growth of data has been a trigger in the emergence of Artificial Intelligence (AI). These data-driven feedback loops enabled continued customization, enabling AI to evolve and representing a new technological leap forward. Nevertheless, many challenges also arise in the field of AI. The first of these challenges is the quality and quantity of the data required to feed the AI model, as effective training requires a significant amount of data, which necessitates large and diverse datasets. Another issue is that data privacy concerns have arisen as AI systems internalize confidential information from fields such as health, finance, aerospace and defense. Having imbalanced classes within datasets is another challenge since it specifically affects the fairness and accuracy of classification algorithms. To overcome these challenges, Data Science (DS) and AI experts are developing various methods of synthesizing and/or augmenting data. Great success has been achieved in the generation of different data types with Generative Adversarial Networks (GANs), which play an important role in these studies. This thesis focuses on improving minority class in tabular data. Using various open source imbalanced class datasets with different volumes, the study leverages various GAN models to augment minority class. The original and augmented datasets are then compared using statistical visualizations and Machine Learning (ML) model performances. In conclusion, the research highlights the important role of GANs in addressing data-related challenges in AI and demonstrates their effectiveness in rebalancing unstable datasets for improved model performance.

Subject Keywords

Data augmentation, Synthetic data generation, Generative adversarial networks, Machine learning, Deep learning

URI

https://hdl.handle.net/11511/105440

Collections

Graduate School of Natural and Applied Sciences, Thesis