Research map / Detection and segmentation / Open-set and few-shot detection
Contradicted in Reliable, Replicated in Unreliable: Dual-Source Reference for Fake News Early Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RUSR, an early fake news detection method that builds a dual-source (reliable and unreliable) publication background and jointly trains support-degree, semantic-outlier-degree and content models.
Problem
Early detection of fake news, where existing methods use reliable-source support but ignore differences in semantic outlier degree of news among unreliable-source information published in the same period.
Approach
Fake news is often accompanied by many unreliable-source articles promoting the same propaganda, so it forms a tighter semantic cluster than real news; RUSR combines reliable-source support evaluation with a two-hop semantic outlier evaluation over a publication background, using expanding-range relative time encoding and a redistribution module within a multi-model joint training and collaborativ
Key results
- RUSR achieves average macF1 0.727 and accuracy 0.734 across Q1-Q4, versus 0.693 for the best baseline (DeClarE+NEP, macF1 0.693 vs Emo 0.713 average macF1 is higher; RUSR best overall)
- Ablation: removing the redistribution module's rotation step drops Modelr macF1 from 0.738 to 0.382 on Q4
- Removing the time encoding lowers average macF1 from 0.727 to 0.709
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Modelc to the RUSR framework (ablation) | RUSR · AVG | macF1: 0.692 -> 0.727 | Helped |
| add Modelr to the RUSR framework (ablation) | RUSR · AVG | macF1: 0.710 -> 0.727 | Helped |
| add Modelu to the RUSR framework (ablation) | RUSR · AVG | macF1: 0.714 -> 0.727 | Helped |
| add relative time encoding (ablation) | RUSR · AVG | macF1: 0.709 -> 0.727 | Helped |
| add the Mdr redistribution module (ablation) | RUSR · AVG | macF1: 0.693 -> 0.727 | Helped |
| add the Mdu redistribution module (ablation) | RUSR · AVG | macF1: 0.712 -> 0.727 | Helped |
| add the rotation step to Mdr (ablation) | RUSR · Q4 | macF1: 0.382 -> 0.738 | Helped |
| add the rotation step to Mdu (ablation) | RUSR · Q4 | macF1: 0.705 -> 0.752 | Helped |
| remove Modelc (ablation) | RUSR joint training · AVG | Modelr macF1: 0.715 -> 0.690 | Hurt |
| remove Modelr (ablation) | RUSR joint training · AVG | Modelc macF1: 0.710 -> 0.721 | Hurt |
| remove Modelu (ablation) | RUSR joint training · AVG | Modelc macF1: 0.721 -> 0.721 | Hurt |
| increase the maximum background size beyond 110 (ablation) | RUSR Modelu · Q1-Q4 | macF1 | Hurt |
Surprising findings
- Methods using evidence or news environment do not necessarily outperform content-only methods, and adding NEP does not always help.
- Without the redistribution module, the similarity distributions collapse to [0.5,1] for both real and fake news, deviating from the expected distribution.
Limitations
- Dataset is English only and relies on Politifact/Snopes labels with a limited sample size (about 5,855 verified news).
- Pre-trained encoders are not fine-tuned due to the small dataset.
- Performance of the outlier model declines when background size grows because unreliable background news outside the detected period increases.
Experimental setup
Datasets: Politifact Snopes NELA-GT-2019-2022 NELA-GT-2020 reliability labels
Benchmarks: RUSR dataset Q1-Q4 2022 splits
Metrics: macro F1 accuracy F1 fake F1 real
Baselines: EANNT Emo BERT+FTT DeClarE MAC Emo+NEP DeClarE+NEP
Builds on: DeClarE MAC EANNT Sheng et al. 2022 news environment perception (NEP) BERT+FTT
Closest papers by meaning
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.88)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.86)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.86)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.84)
- Harnessing Network Effect for Fake News Mitigation: Selecting Debunkers via Self-Imitation Learning (AAAI 2024 · similarity 0.84)
- External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.