Source-linked AI summary
Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks
Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, Junzhou Huang
TL;DR
Rumor detection needs representations of both propagation and dispersion because existing handcrafted and graph-based approaches have important limitations. Bi-GCN addresses this gap with top-down and bottom-up graph convolutions while repeatedly incorporating source-post features. Across three real-world datasets, the GCN-based approaches outperform state-of-the-art baselines in accuracy and efficiency, with Bi-GCN achieving the best performance.
Problem
Existing methods either rely on labor-intensive handcrafted features or overlook propagation direction and wide dispersion, despite both being important rumor characteristics.
Method
Bi-GCN uses TD-GCN to model top-down rumor propagation, BU-GCN to model bottom-up rumor dispersion, and concatenates source-post features at each graph-convolutional layer.
Results
Across three real-world datasets, GCN-based approaches outperform state-of-the-art baselines in accuracy and efficiency, while Bi-GCN achieves the best performance and stronger early-detection effectiveness.
Takeaways & Limitations
The study supports using directed, bidirectional graph modeling with root-feature enhancement for rumor detection on social media.
Takeaways & Limitations
The paper’s motivation identifies handcrafted approaches as limited by time-consuming feature engineering and insufficient high-level representations of propagation and dispersion.
Abstract
from arXiv · showhide
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.
Introduction
Rumor detection must address both deep propagation and wide dispersion, which conventional and sequential deep-learning approaches do not jointly capture. Bi-GCN uses directed graph convolutions in both directions, incorporates source-post features at every layer, and outperforms prior approaches, including for early detection.
- Rumors spread widely and quickly on social media, creating societal harm and economic losses that motivate efficient, early identification.
- Handcrafted-feature methods are time-consuming and labor-intensive, while their features often lack high-level representations of rumor propagation and dispersion.
- Sequential deep-learning models learn temporal propagation features but neglect the wider dispersion structures involved in rumor detection.
- Undirected GCN aggregates relationship information but loses follow-order information and does not consider propagation direction.
- Bi-GCN combines TD-GCN for top-down propagation with BU-GCN for bottom-up dispersion, then merges their learned representations.
- Concatenating source-post features at every graph-convolutional layer strengthens use of root information, while experiments show higher effectiveness for early rumor detection.
Related Work
Prior rumor-detection work uses handcrafted features, temporal or structural modeling, recurrent networks, and convolutional approaches. GCNs offer stronger global graph-structure modeling, motivating their application to rumor propagation trees.
- Earlier methods classify rumors using text contents, user profiles, propagation structures, and other handcrafted social-context features.
- Figure 1 contrasts undirected node relationships with top-down propagation and bottom-up aggregation of wide dispersion.
- Deep-learning approaches use recurrent models to capture hidden representations from temporal content and attention mechanisms to focus on text features.
- GCN can capture global structural features from graphs or trees more effectively than the deep-learning models discussed in the related work.
- Graph-convolution research developed methods for both undirected and directed graphs, including spectral filters and first-order approximations.
Preliminaries
The paper defines rumor events as source posts, responsive posts, propagation structures, features, and labels, then describes GCN message passing and DropEdge regularization.
- Event and graph notation: Each event contains a source post, responsive posts, a propagation graph, a feature matrix, and a ground-truth rumor label.Labels may be binary False/True Rumor or four-class Non-rumor, False Rumor, True Rumor, and Unverified Rumor.
- Event and graph notation: The propagation graph is directed from responded posts toward retweeted or responsive posts, with the source post as its root.Edges encode response and retweet relationships between posts.
- Event and graph notation: Rumor detection learns a classifier that predicts an event label from text, user information, and propagation structure.The event representation includes the related posts and their constructed propagation graph.
- Graph convolutional networks: GCN message passing computes each layer’s hidden feature matrix from the adjacency matrix, previous hidden features, and trainable parameters.The message propagation function M operates on graph structure and learned node representations.
- Graph convolutional networks: The first-order ChebNet formulation adds self-connections, normalizes the adjacency matrix, applies trainable weights, and uses an activation function such as ReLU.The normalized adjacency is based on the degree matrix of the self-connected graph.
- Regularization: DropEdge randomly removes a proportion of graph edges during each training epoch to create varied graph copies and reduce over-fitting.With Ne edges and dropping rate p, Adrop is formed from Ne × p randomly sampled original edges.
Bi-GCN Rumor Detection Model
Bi-GCN is an effective GCN-based rumor detector that learns representations from both rumor propagation and dispersion using bidirectional graph processing.
- Model overview: Bi-GCN uses two-layer first-order ChebNet components to learn high-level representations from rumor propagation and rumor dispersion.The model is organized as a four-step rumor detection process.
- Model overview: The model applies top-down and bottom-up graph convolutional networks to capture complementary rumor structures.TD-GCN and BU-GCN form the two directional components of Bi-GCN.
- Model overview: Bi-GCN processes one event at a time, mapping an event ci to its predicted label yi.The paper omits the event subscript in subsequent model descriptions for readability.
- Model overview: Figure 2 distinguishes the original feature matrix X, layer-specific hidden features Hk, and source-post features used in the model.The caption identifies Hk as the hidden matrix generated by the k-th graph convolutional layer.
1 Construct Propagation and Dispersion Graphs
The model constructs a directed rumor propagation graph from retweet and response relationships, then applies edge dropping before directional representation learning.
- Graph construction: For each rumor event, the propagation structure is represented by an adjacency matrix A and feature matrix X derived from its spreading tree.A contains edges from upper nodes to lower nodes in the rumor tree.
- Graph construction: At each training epoch, a percentage p of edges is dropped from A to form A′ and help avoid potential overfitting.The operation follows the DropEdge formulation.
2 Calculate the High-level Node Representations
After DropEdge, TD-GCN and BU-GCN compute complementary high-level node representations from top-down propagation and bottom-up dispersion structures.
- Directional representation learning: DropEdge precedes directional encoding, producing top-down propagation features and bottom-up dispersion features through TD-GCN and BU-GCN.The two networks operate on opposite directions of the rumor structure.
- Directional representation learning: TD-GCN applies the GCN message-passing equation over two layers using the top-down adjacency matrix and input features.The resulting hidden matrices represent the two TD-GCN layers.
- Directional representation learning: TD-GCN uses trainable filter matrices, ReLU activation, and dropout on graph convolutional layers to reduce over-fitting.BU-GCN computes its hidden features analogously to TD-GCN.
- Directional representation learning: BU-GCN calculates bottom-up hidden features in the same manner as the two-layer TD-GCN equations.This creates a counterpart representation for the reverse graph direction.
3 Root Feature Enhancement
Root feature enhancement incorporates the source post’s hidden representation into each graph-convolution layer, strengthening node representations for rumor detection. The operation is applied to both TD-GCN and BU-GCN.
- The source post contains abundant information and can broadly influence a rumor event, motivating representations that model relationships between nodes and the source post.
- Root feature enhancement concatenates every node’s hidden feature with the root node’s hidden feature from the preceding graph-convolution layer.
- The enhanced TD-GCN replaces the preceding hidden-feature representation with a new feature matrix formed using the root feature.
- BU-GCN obtains its root-enhanced hidden features in the same manner as TD-GCN.
4 Representations of Propagation and Dispersion for Rumor Classification
Bi-GCN aggregates node representations from TD-GCN and BU-GCN to encode rumor propagation and dispersion, then combines both representations for event classification. The final prediction uses fully connected and softmax layers and is trained with regularized cross-entropy.
- Mean pooling aggregates TD-GCN and BU-GCN node representations into propagation and dispersion representations, respectively.
- The model concatenates the propagation and dispersion representations to merge information from both graph directions.
- Fully connected layers followed by softmax produce probabilities for all event classes used to predict the event label.
- Training minimizes cross-entropy between predictions and ground-truth distributions while applying an L2 regularizer to model parameters.
Experiments
Experiments evaluate Bi-GCN on three real-world datasets against handcrafted-feature, recurrent, recursive, and graph-based baselines. Bi-GCN outperforms these approaches, while ablations and early-detection tests support combining directional structures with source-post information.
- Experimental Setup: Experiments use Weibo, Twitter15, and Twitter16, with user nodes, retweet or response edges, and TF-IDF features.Weibo has false- and true-rumor labels; Twitter15 and Twitter16 use four finer-grained classes.
- Overall Performance: Deep learning baselines outperform handcrafted-feature baselines, while Bi-GCN surpasses PPC RNN+CNN on all performance measures.The authors attribute the latter comparison to incorporating rumor-dispersion structure through graph processing.
- Overall Performance: Bi-GCN is significantly superior to RvNN, and root feature enhancement directs greater attention to source-post information.The authors connect this design to RvNN’s reliance on leaf-node representations and the limited information in latest posts.
- Ablation Study: Adding root feature enhancement improves every evaluated GCN variant, and Bi-GCN is always superior to UD-GCN, TD-GCN, and BU-GCN.The ablation compares directed and undirected variants with and without concatenated root features.
- Early Rumor Detection: Bi-GCN reaches relatively high accuracy early and is remarkably superior to other compared models at every detection deadline.Early detection uses only posts released before successive deadlines; PPC RNN+CNN is excluded because it cannot process varying-length data accurately.
Conclusions
Bi-GCN combines graph-based modeling of rumor propagation and dispersion, with source-post features injected throughout the network. Across three real-world datasets, it outperforms state-of-the-art baselines in accuracy and efficiency, achieving the best performance among its variants.
- Bi-GCN models both top-down propagation along relationship chains and bottom-up dispersion within communities.It also concatenates source-post features after each graph convolutional layer.
- The proposed GCN-based approaches outperform state-of-the-art baselines by large margins in both accuracy and efficiency.
- Bi-GCN achieves the best performance among the proposed variants by considering propagation and dispersion together.The variants include UD-GCN, TD-GCN, and BU-GCN.