Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive Token-level Explanations for Graph-based Rumour Detection

Daniel Wai Kit Chin, Roy Ka-Wei Lee

TL;DR

This work tackles the opacity of GNN-based rumour detection by introducing CT-LRP, a token-level, contrastive explanation framework that backpropagates relevance from the predicted class through both the graph and text embedding components. By decomposing explanations into class-specific token attributions, CT-LRP captures fine-grained dependencies in high-dimensional text embeddings and provides a robust mechanism to disambiguate tokens that influence multiple classes. The authors extend evaluation metrics to token-level resolution and demonstrate through experiments on three public datasets and three representative GNNs that CT-LRP achieves higher fidelity and a favorable sparsity balance compared with baselines. This token-focused explainability advances trust and interpretability in misinformation detection, with potential applicability to related tasks such as fake news and multimodal graphs.

Abstract

The widespread use of social media has accelerated the dissemination of information, but it has also facilitated the spread of harmful rumours, which can disrupt economies, influence political outcomes, and exacerbate public health crises, such as the COVID-19 pandemic. While Graph Neural Network (GNN)-based approaches have shown significant promise in automated rumour detection, they often lack transparency, making their predictions difficult to interpret. Existing graph explainability techniques fall short in addressing the unique challenges posed by the dependencies among feature dimensions in high-dimensional text embeddings used in GNN-based models. In this paper, we introduce Contrastive Token Layerwise Relevance Propagation (CT-LRP), a novel framework designed to enhance the explainability of GNN-based rumour detection. CT-LRP extends current graph explainability methods by providing token-level explanations that offer greater granularity and interpretability. We evaluate the effectiveness of CT-LRP across multiple GNN models trained on three publicly available rumour detection datasets, demonstrating that it consistently produces high-fidelity, meaningful explanations, paving the way for more robust and trustworthy rumour detection systems.

Contrastive Token-level Explanations for Graph-based Rumour Detection

TL;DR

This work tackles the opacity of GNN-based rumour detection by introducing CT-LRP, a token-level, contrastive explanation framework that backpropagates relevance from the predicted class through both the graph and text embedding components. By decomposing explanations into class-specific token attributions, CT-LRP captures fine-grained dependencies in high-dimensional text embeddings and provides a robust mechanism to disambiguate tokens that influence multiple classes. The authors extend evaluation metrics to token-level resolution and demonstrate through experiments on three public datasets and three representative GNNs that CT-LRP achieves higher fidelity and a favorable sparsity balance compared with baselines. This token-focused explainability advances trust and interpretability in misinformation detection, with potential applicability to related tasks such as fake news and multimodal graphs.

Abstract

The widespread use of social media has accelerated the dissemination of information, but it has also facilitated the spread of harmful rumours, which can disrupt economies, influence political outcomes, and exacerbate public health crises, such as the COVID-19 pandemic. While Graph Neural Network (GNN)-based approaches have shown significant promise in automated rumour detection, they often lack transparency, making their predictions difficult to interpret. Existing graph explainability techniques fall short in addressing the unique challenges posed by the dependencies among feature dimensions in high-dimensional text embeddings used in GNN-based models. In this paper, we introduce Contrastive Token Layerwise Relevance Propagation (CT-LRP), a novel framework designed to enhance the explainability of GNN-based rumour detection. CT-LRP extends current graph explainability methods by providing token-level explanations that offer greater granularity and interpretability. We evaluate the effectiveness of CT-LRP across multiple GNN models trained on three publicly available rumour detection datasets, demonstrating that it consistently produces high-fidelity, meaningful explanations, paving the way for more robust and trustworthy rumour detection systems.

Paper Structure

This paper contains 23 sections, 6 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. GNN Explainability
  4. Text Explainability
  5. Preliminaries
  6. Proposed Framework
  7. Task Formulation
  8. GNN explanation
  9. Token-level explanation
  10. Contrastive Token-Level Explanation
  11. Experiments
  12. Models and Datasets
  13. Models
  14. Datasets
  15. Data Preprocessing and Model Training
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Overview of the proposed CT-LRP framework showing the flow of information through the GNN and text embedding function. Inputs to the forward and backward pass are colour-coded in green, intermediate outputs in purple and final outputs in blue.
  • Figure 2: Token-level Explanation produced by CT-LRP. Node-level attribution is shown with darker shades of blue indicating greater attribution from that node. Token-level explanation with blue and green highlights for class-specific and common task-relevant attribution, and red highlights for negative attribution. Darker shades indicate a greater magnitude of importance. The input is the source claim and the responses are modelled as a graph.
  • Figure 3: Quantitative performance comparisons on three models trained on three datasets with the curves obtained by varying sparsity levels.