Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Two Birds with One Stone: Improving Rumor Detection by Addressing the Unfairness Issue

Junyi Chen, Mengjia Wu, Qian Liu, Ying Ding, Yi Zhang

TL;DR

Rumor detectors are prone to non-causal shortcuts from confounding sensitive attributes, hurting both accuracy and fairness. The paper introduces a plug-and-play two-step framework that first identifies worst-case partitions via dynamic unfair partitioning and then trains invariant representations through constrained learning, all without annotated sensitive attributes. Empirical results on FineFake and PHEME show consistent gains in detection performance and fairness across multiple base detectors, with ablations and visualizations supporting the importance of dynamic partitions, supervised contrastive loss, and the variance term. This approach enables more robust and fair rumor detection across diverse, real-world groupings and lays groundwork for fair benchmarks and sparse attribute integration.

Abstract

The degraded performance and group unfairness caused by confounding sensitive attributes in rumor detection remains relatively unexplored. To address this, we propose a two-step framework. Initially, it identifies confounding sensitive attributes that limit rumor detection performance and cause unfairness across groups. Subsequently, we aim to learn equally informative representations through invariant learning. Our method considers diverse sets of groups without sensitive attribute annotations. Experiments show our method easily integrates with existing rumor detectors, significantly improving both their detection performance and fairness.

Two Birds with One Stone: Improving Rumor Detection by Addressing the Unfairness Issue

TL;DR

Rumor detectors are prone to non-causal shortcuts from confounding sensitive attributes, hurting both accuracy and fairness. The paper introduces a plug-and-play two-step framework that first identifies worst-case partitions via dynamic unfair partitioning and then trains invariant representations through constrained learning, all without annotated sensitive attributes. Empirical results on FineFake and PHEME show consistent gains in detection performance and fairness across multiple base detectors, with ablations and visualizations supporting the importance of dynamic partitions, supervised contrastive loss, and the variance term. This approach enables more robust and fair rumor detection across diverse, real-world groupings and lays groundwork for fair benchmarks and sparse attribute integration.

Abstract

The degraded performance and group unfairness caused by confounding sensitive attributes in rumor detection remains relatively unexplored. To address this, we propose a two-step framework. Initially, it identifies confounding sensitive attributes that limit rumor detection performance and cause unfairness across groups. Subsequently, we aim to learn equally informative representations through invariant learning. Our method considers diverse sets of groups without sensitive attribute annotations. Experiments show our method easily integrates with existing rumor detectors, significantly improving both their detection performance and fairness.
Paper Structure (11 sections, 3 equations, 3 figures, 2 tables)

This paper contains 11 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Unfair Partition
  4. Invariant Learning
  5. Experimental Evaluation
  6. Setup
  7. Result
  8. Performance Comparison
  9. Ablative Study
  10. Visual Evidence and Effective Intervention
  11. Conclusion

Figures (3)

  • Figure 1: Schematic view of our proposed method.
  • Figure 2: Visualization of learned features. Left: MDFEND. Right: MDFEND w/ our method. Colors denote different groups. Fewer points appear visible with our method due to overlap.
  • Figure 3: Intervention effects of our method on the base model. W->C: proportion of base model's wrong predictions corrected by our method; C->W: proportion of correct predictions made incorrect.