Abstract:The rapid spread of misinformation on social media, especially during crises, challenges public decision-making. To address this, we propose HierTKG, a framework combining Temporal Graph Networks (TGN) and hierarchical pooling (DiffPool) to model rumor dynamics across temporal and structural scales. HierTKG captures key propagation phases, enabling improved temporal link prediction and actionable insights for misinformation control. Experiments demonstrate its effectiveness, achieving an MRR of 0.9845 on ICEWS14 and 0.9312 on WikiData, with competitive performance on noisy datasets like PHEME (MRR: 0.8802). By modeling structured event sequences and dynamic social interactions, HierTKG adapts to diverse propagation patterns, offering a scalable and robust solution for real-time analysis and prediction of rumor spread, aiding proactive intervention strategies.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the real - time prediction and control of rumor spreading on social media platforms. Specifically, the author aims to develop a framework - HierTKG, which can capture the temporal dynamics and structural hierarchical features of rumor spreading by combining Temporal Graph Networks (TGN) and hierarchical pooling techniques (such as DiffPool). The goal of this framework is to improve the understanding of rumor - spreading patterns, thereby providing actionable insights for rumor control.
### Problem Background
1. **Challenges of Rumor Spreading**:
- Information spreads rapidly on social media, especially during crises, and the spread of misinformation may affect public decision - making.
- Rumor - spreading patterns change over time and may peak at specific moments due to influential users or external events.
2. **Limitations of Existing Methods**:
- Existing methods lack a comprehensive framework to simultaneously combine temporal and hierarchical perspectives, limiting their ability to capture the real - time dynamic evolution of rumors.
### Paper Objectives
1. **Model the Temporal Dynamics and Hierarchical Structure of Rumor Spreading**:
- By integrating TGN and hierarchical pooling techniques, model the evolution of interactions between entities (such as users, posts) over time.
- Capture key spreading stages, making improved temporal link prediction possible, thereby providing viable insights regarding rumor control.
2. **Improve the Accuracy of Temporal Link Prediction**:
- By modeling multi - scale structures and temporal dynamics, enhance the predictive ability for rumor spreading and other temporal link prediction tasks.
3. **Provide a Comprehensive Evaluation**:
- Through extensive experiments (such as ablation studies and comparisons with the latest methods), verify the effectiveness and robustness of the proposed framework.
### Main Contributions
1. **Temporal Knowledge Graph Construction**:
- Construct a comprehensive temporal knowledge graph from the PHEME dataset, capturing the detailed temporal and relational information required for rumor spreading.
2. **HierTKG Framework**:
- Introduce a new method of integrating TGN and hierarchical pooling (DiffPool) for modeling knowledge graphs of temporal evolution and hierarchical structure, providing detailed insights into the dynamics of rumor spreading.
3. **Enhanced Temporal Link Prediction**:
- By integrating multi - scale structures and temporal dynamics, improve the accuracy of predicting rumor spreading.
4. **Comprehensive Evaluation**:
- Extensive experiments, including ablation studies, demonstrate the effectiveness and robustness of the framework on different datasets.
### Mathematical Formulas
To ensure the correctness and readability of the formulas, the following are the key formulas involved in the paper:
- **Scoring Function**:
\[
\hat{y}(s, o, t) = f(x_s(t), x_o(t), r, t)
\]
where \(\hat{y}(s, o, t)\in[0, 1]\) represents the predicted link probability.
- **Loss Function**:
\[
L = -\sum_{(s, o, t)\in D^+}\log\hat{y}(s, o, t)-\sum_{(s, o, t)\in D^-}\log(1 - \hat{y}(s, o, t))
\]
- **Node Embedding Update**:
\[
m_v(t)=\text{GatingFunction}(m_v(t - 1),\text{NewMessage})
\]
- **Hierarchical Embedding**:
\[
S^{(l)}=\text{softmax}(W_s^{(l)}\cdot H^{(l - 1)}(t))
\]
\[
H^{(l)}(t)=S^{(l)^\top}H^{(l - 1)}(t)
\]
\[
A^{(l)}=S^{(l)^\top}A^{(l - 1)}S^{(l)}
\]
Through these formulas, the HierTKG framework can effectively capture the temporal and hierarchical features of rumor spreading, thereby improving the accuracy and reliability of prediction.