Scholarlink

<p>As social media usage continues to grow across the globe, the detection of fake accounts has become increasingly challenging and crucial for platform integrity. A major obstacle in this domain lies in the complexity of graph-structured data, which contains numerous nodes and unrelated features, often making data analysis and its manipulation harder. Additionally, such data structures frequently suffer from imbalanced classes, which creates further challenges for accurate detection. For our study, we used Twitter’s (now “X”) social media data, represented as a graph structure, where nodes represent individual users, and edges denote the connections between them, forming a vast and intricate network. This approach of graph representation can effectively address significant issues related to both imbalanced datasets and limited labeled data. Therefore, we implemented an Autoencoder along with a type of neural network technique called a Semi-supervised Generative Adversarial Network (SGAN), which can handle scenarios with few labeled samples. The results of this test ultimately showed that the accuracy reached 81% in detecting fake accounts, even when using only 100 labeled samples.</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

Fake detection in imbalance dataset by semi-supervised learning with GAN

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Jinus Bordbar,
  • Saman Ardalan,
  • Mohammadreza Mohammadrezaei,
  • Zahra Ghasemi

摘要

As social media usage continues to grow across the globe, the detection of fake accounts has become increasingly challenging and crucial for platform integrity. A major obstacle in this domain lies in the complexity of graph-structured data, which contains numerous nodes and unrelated features, often making data analysis and its manipulation harder. Additionally, such data structures frequently suffer from imbalanced classes, which creates further challenges for accurate detection. For our study, we used Twitter’s (now “X”) social media data, represented as a graph structure, where nodes represent individual users, and edges denote the connections between them, forming a vast and intricate network. This approach of graph representation can effectively address significant issues related to both imbalanced datasets and limited labeled data. Therefore, we implemented an Autoencoder along with a type of neural network technique called a Semi-supervised Generative Adversarial Network (SGAN), which can handle scenarios with few labeled samples. The results of this test ultimately showed that the accuracy reached 81% in detecting fake accounts, even when using only 100 labeled samples.