Good Advisor for Source Localization: Using Large Language Model to Guide the Source Inference Process

Dongpeng Hou, Wenfei Wei, Chao Gao, Xianghua Li, Zhen WangIJCAI 2025 · DOI

Abstract

Identifying the original source of a rumor on social media is a high-stakes "needle in a haystack" problem. Traditional methods often suffer from either a "semantic gap"—ignoring the rich textual content of comments—or a "structural gap"—failing to grasp complex network propagation patterns. This paper introduces CRSLL (Contrastive Rumor Source Localization via LLM), a framework that utilizes Large Language Models (LLMs) as "expert advisors." By fusing deep semantic analysis of user comments with graph-based propagation dynamics, CRSLL locates rumor sources with unprecedented accuracy, significantly outperforming both traditional graph neural networks and direct LLM prompting.

Introduction

When a rumor goes viral, finding "patient zero" is essential for effective mitigation. Historically, researchers have relied on graph-based methods that analyze network snapshots or LLM-based methods that process text. However, graph models often miss human intent, while LLMs struggle with high-dimensional graph topologies.

Comparison of direct LLM prompting vs. the proposed guided framework.

The authors identify a fundamental mismatch: LLMs are brilliant at reasoning about social context but poor at calculating network positions. CRSLL addresses this by shifting the LLM's role from a direct decoder to an advisor that provides intermediate semantic reasoning, bridging the gap between linguistic nuance and structural mechanics.

Methodology: The CRSLL Framework

The innovation of CRSLL lies in its multi-modal fusion approach, treating the LLM as a source of expert analysis rather than a final decision-maker.

The CRSLL architecture showing multi-modal fusion and contrastive learning components.

LLM as Advisor

Instead of asking the LLM to pick the source directly, CRSLL uses prompt engineering to generate a "Source Analysis." This analysis explains why a specific comment might indicate a user is a source (e.g., initiating a claim vs. reacting to one). This reasoning is then converted into embeddings to serve as a high-level feature for the inference engine.

Contrastive Learning and Feature Masking

To sharpen the model's focus, CRSLL employs contrastive learning to distinguish between comments in rumor cascades and non-rumor cascades. This helps the model identify specific "rumor-like" linguistic patterns. Additionally, a differentiable masking strategy using Gumbel-Softmax is applied to prune redundant or noisy features from high-dimensional text embeddings, ensuring only discriminative data is used.

Self-Loop Attention GCN

Standard Graph Convolutional Networks (GCNs) often struggle with "celebrity" nodes—users with massive follower counts whose influence can dilute information during aggregation. CRSLL introduces a learnable self-loop attention mechanism to better weight these influential nodes, acknowledging that a user's own profile is often more indicative of their role than the average of their neighbors.

Experimental Results

The framework was evaluated on three real-world datasets: Twitter15, Twitter16, and Weibo. The authors expanded these datasets to include raw comments and detailed user profiles.

CRSLL demonstrated a significant performance leap:

  • Superiority over GPT-4: CRSLL improved upon direct GPT-4 prompting by an average of 62.3% across datasets.
  • High Precision: The model achieved F1-scores of 0.951 on Twitter15 and 0.946 on Twitter16, whereas traditional GNNs and direct LLM baselines hovered around 0.55-0.60.
  • Ablation Success: Removing the LLM-generated "advice" led to a substantial drop in performance, confirming that semantic reasoning is a critical component of modern source localization.

Conclusion and Future Work

CRSLL represents a paradigm shift in how LLMs are integrated into graph-based tasks. By leveraging LLMs for reasoning rather than calculation, the framework overcomes the structural limitations of language models. While the computational cost of processing every comment through an LLM remains a challenge for real-time scaling, the "advisor" model provides a robust blueprint for future multi-modal social network analysis. Future work will explore more efficient integration methods and apply this framework to other complex tasks like community detection.

Summary updated

BibTeX
@inproceedings{hou2025good,
  author = {Dongpeng Hou and Wenfei Wei and Chao Gao and Xianghua Li and Zhen Wang},
  title = {{Good Advisor for Source Localization: Using Large Language Model to Guide the Source Inference Process}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/326},
}
Figure