[Paper Review] Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks

Tian Bian, Xi Xiao+5 more|arXiv (Cornell University)|Jan 17, 2020

Misinformation and Its Impacts28 references21 citations

TL;DR

This paper proposes Bi-Directional Graph Convolutional Networks (Bi-GCN), a novel GCN-based model that jointly captures top-down rumor propagation and bottom-up dispersion patterns in social media by leveraging dual graph convolutions on directed rumor trees. By integrating source post features at each layer and using DropEdge for regularization, Bi-GCN achieves state-of-the-art performance in both rumor detection and early detection across three real-world datasets.

ABSTRACT

Social media has been developing rapidly in public due to its nature of spreading new information, which leads to rumors being circulated. Meanwhile, detecting rumors from such massive information in social media is becoming an arduous challenge. Therefore, some deep learning methods are applied to discover rumors through the way they spread, such as Recursive Neural Network (RvNN) and so on. However, these deep learning methods only take into account the patterns of deep propagation but ignore the structures of wide dispersion in rumor detection. Actually, propagation and dispersion are two crucial characteristics of rumors. In this paper, we propose a novel bi-directional graph model, named Bi-Directional Graph Convolutional Networks (Bi-GCN), to explore both characteristics by operating on both top-down and bottom-up propagation of rumors. It leverages a GCN with a top-down directed graph of rumor spreading to learn the patterns of rumor propagation, and a GCN with an opposite directed graph of rumor diffusion to capture the structures of rumor dispersion. Moreover, the information from the source post is involved in each layer of GCN to enhance the influences from the roots of rumors. Encouraging empirical results on several benchmarks confirm the superiority of the proposed method over the state-of-the-art approaches.

Motivation & Objective

• Address the limitations of existing rumor detection methods that rely on handcrafted features or sequential models like RvNN, which fail to model global structural patterns in rumor spread.
• Overcome the shortcomings of undirected GCNs (UD-GCN) that lose propagation directionality and of CNNs that cannot model global graph structures.
• Develop a unified deep learning framework that simultaneously models top-down propagation and bottom-up dispersion in rumor propagation trees.
• Improve early rumor detection by leveraging structural representations from both propagation and dispersion patterns.
• Enhance model robustness and performance by integrating source post features at every GCN layer and applying DropEdge regularization.

Proposed method

• Construct two directed graphs from rumor propagation trees: a top-down (TD) graph for propagation and a bottom-up (BU) graph for dispersion.
• Apply graph convolutional networks (GCNs) on both TD and BU graphs to learn high-level representations of rumor propagation and structural dispersion.
• Concatenate the source post’s features with hidden representations at each GCN layer to amplify root influence and improve feature learning.
• Merge the final representations from TD-GCN and BU-GCN via fully connected layers for end-to-end classification.
• Apply DropEdge during training to prevent overfitting by randomly removing edges from the graph, improving generalization.
• Use a multi-layer GCN architecture with ReLU activation and residual connections to capture hierarchical patterns in rumor structures.

Experimental results

Research questions

• RQ1 Can a GCN-based model effectively capture both the causal propagation patterns and structural dispersion patterns in rumor spread?
• RQ2 Does integrating source post features at every GCN layer significantly improve rumor detection performance compared to standard GCN or RvNN?
• RQ3 How does the bi-directional GCN (Bi-GCN) compare to unidirectional GCNs (TD-GCN, BU-GCN) and undirected GCNs (UD-GCN) in terms of detection accuracy?
• RQ4 Can the proposed model achieve superior early rumor detection performance by leveraging structural features from both propagation and dispersion?
• RQ5 To what extent does DropEdge improve generalization and reduce overfitting in rumor detection with graph-structured data?

Key findings

• Bi-GCN achieves state-of-the-art performance on three real-world datasets (Weibo, Twitter15, Twitter16), outperforming all baseline methods including RvNN, SVM, and DTC in both accuracy and F1-score.
• The ablation study confirms that concatenating source post features at each GCN layer significantly improves performance, with the best results achieved when root features are integrated throughout the network.
• Bi-GCN consistently outperforms UD-GCN, TD-GCN, and BU-GCN, demonstrating the necessity of modeling both top-down propagation and bottom-up dispersion simultaneously.
• In early rumor detection, Bi-GCN achieves high accuracy even at very early deadlines (e.g., within 1–2 hours of source post), significantly outperforming RvNN and other baselines.
• The model’s performance remains robust across datasets, with the highest F1-score of 0.872 on the Weibo dataset and 0.821 on Twitter16, indicating strong generalization.
• DropEdge effectively reduces overfitting, as shown by improved validation performance and consistent gains across all datasets and evaluation metrics.