Comparative Analysis of Graph Neural Network with SAGE Conv, GAT Conv, and GCN Conv Techniques for Fake News Detection

Abstract

In today's rapidly evolving information landscape, the spread of fake news poses a critical challenge to the integrity of public information. Fake news, characterized by intentionally falsified or misrepresented information, can manipulate public opinion, disrupt political processes, and incite social instability. Consequently, the detection of fake news has become essential for maintaining media integrity and ensuring a healthy democratic function. Traditional methods for detecting fake news, such as decision trees and support vector machines, often fall short due to their inability to capture the relational and structural context of data. To address this, Graph Neural Networks (GNNs) have emerged as promising solutions, offering the ability to process data structured as graphs and retain topological information. This study investigates three GNN models—Graph Convolutional Network (GCN Conv), Graph Attention Network (GAT Conv), and GraphSAGE (SAGEConv)—each with unique strategies for handling graph data in the context of fake news detection. Our comparative analysis reveals that GAT Conv achieves the highest test accuracy of 0.9488 at epoch 86, demonstrating strong learning performance and efficient convergence. SAGE Conv, while slightly less effective, achieves a maximum accuracy of 0.9472 at epoch 93, indicating its potential in specific scenarios. GCN Conv offers a balanced performance with a maximum accuracy of 0.9482 at epoch 99, showcasing its robustness as an alternative approach. These findings underscore the importance of selecting suitable GNN models based on the characteristics of the network, optimizing fake news detection efforts, and contributing to enhanced media integrity and democratic stability.

Keywords: GNN, Fake News Detection, Deep Learning