Research map / Reinforcement learning / Imitation learning
Harnessing Network Effect for Fake News Mitigation: Selecting Debunkers via Self-Imitation Learning
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 NAGASIL, a generative adversarial self-imitation learning method with negative samples and augmented states for selecting debunkers to counter fake news under episodic rewards on SEIR-simulated social networks.
Problem
Fake news mitigation via debunker selection is a multi-stage RL problem with episodic rewards (net effect of each debunker is not observable), and existing self-imitation methods are sample-inefficient and assume full state observability.
Approach
Extends GASIL with (i) a negative-sample regularizer that penalises actions favoured by past bad episodes, and (ii) an augmented state that appends discounted history of previous state-action pairs in the same campaign, with theoretical propositions motivating both.
Key results
- NAGASIL outperforms GASIL, SIDP, DQN-FSP, MAX-INF, MAX-DEF and RND across budgets and stage lengths on Facebook and synthetic Twitter networks, per the paper's figures.
- Ablation: NGASIL (negative samples only) clearly beats GASIL; AGASIL (augmented state only) also beats GASIL; full NAGASIL is best.
- NAGASIL keeps its advantage on a larger 2500-user Twitter network (3 runs, box plot).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NAGASIL instead of the reported mitigation baselines | Fake news mitigation policies · Facebook ego networks and synthetic Twitter networks | episodic reward | Helped |
| Use NAGASIL instead of the reported mitigation baselines | Fake news mitigation policies · Twitter networks with varying stage length, density, and size | episodic reward | Helped |
| Add negative samples to GASIL (ablation) | GASIL · Synthetic Twitter networks, ablation study | episodic reward | Helped |
| Add augmented state to GASIL (ablation) | GASIL · Synthetic Twitter networks, ablation study | episodic reward | Helped |
| Add negative samples to the augmented-state variant (ablation) | AGASIL · Synthetic Twitter networks, ablation study | episodic reward | Helped |
| Use augmented state together with negative samples rather than either component alone (ablation) | NAGASIL · Synthetic Twitter networks, ablation study | episodic reward | Helped |
Surprising findings
- PPO and DQN performed consistently worse than random selection (RND), so they were not reported.
- MAX-INF performance varied strongly between the two Facebook ego networks, suggesting sensitivity to network structure.
Limitations
- Assumes the truth value of news is established and provided, so a fake news detector must be used alongside.
- Evaluated only with one information propagation model (SEIR); robustness to other models is future work.
- Experiments use simulated networks with PHEME-derived propagation rather than real deployment.
Experimental setup
Datasets: PHEME Facebook ego networks (SNAP) synthetic Twitter-like scale-free networks
Benchmarks: Facebook ego networks #1 and #2 (SEIR simulation with PHEME propagation) synthetic Twitter networks (1250 and 2500 users)
Metrics: episodic reward (negative log fraction of infected users)
Baselines: DQN-FSP MAX-INF MAX-DEF GASIL SIDP RND GASIL ablations (NGASIL, AGASIL)
Builds on: GASIL GAIL self-imitation learning SIDP
Closest papers by meaning
- Enhancing Robustness of Graph Neural Networks on Social Media with Explainable Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Contradicted in Reliable, Replicated in Unreliable: Dual-Source Reference for Fake News Early Detection (AAAI 2025 · similarity 0.84)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.84)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.83)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.83)
- Unveiling Implicit Deceptive Patterns in Multi-Modal Fake News via Neuro-Symbolic Reasoning (AAAI 2024 · similarity 0.83)
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.