Scholarlink

<p>Handling data imbalance remains a challenge in machine learning, as existing oversampling tactics often face issues like overfitting and high computational cost. Recent tactics leveraging GANs have improved synthetic data quality but still lack efficiency and robustness. In this study, we propose GTGAN, a game-theoretic GAN framework that employs parallel training and ensemble voting based on the deep convolutional network to generate high-quality synthetic minority data efficiently. Unlike traditional oversampling tactics, GTGAN utilizes a competitive learning approach to generate high-quality synthetic data, enhancing the representation of minority groups while minimizing the risk of overfitting. Empirical outcomes on benchmark databases reveal that GTGAN surpasses traditional oversampling tactics, including SMOTE variants, in classification accuracy. Statistical analysis confirms the superior performance of our approach, highlighting its potential for imbalanced data categorization tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GTGAN: A New Framework Based on Game Theory and Generative Adversarial Network for Imbalanced Data Classification

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Yan He,
  • Guozhu Yin

摘要

Handling data imbalance remains a challenge in machine learning, as existing oversampling tactics often face issues like overfitting and high computational cost. Recent tactics leveraging GANs have improved synthetic data quality but still lack efficiency and robustness. In this study, we propose GTGAN, a game-theoretic GAN framework that employs parallel training and ensemble voting based on the deep convolutional network to generate high-quality synthetic minority data efficiently. Unlike traditional oversampling tactics, GTGAN utilizes a competitive learning approach to generate high-quality synthetic data, enhancing the representation of minority groups while minimizing the risk of overfitting. Empirical outcomes on benchmark databases reveal that GTGAN surpasses traditional oversampling tactics, including SMOTE variants, in classification accuracy. Statistical analysis confirms the superior performance of our approach, highlighting its potential for imbalanced data categorization tasks.