Research map / Vision-language and multimodal / Human and embodied modeling
Reinforced Adaptive Knowledge Learning for Multimodal Fake News 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 AKA-Fake, which uses RL to learn a per-news adaptive knowledge subgraph from a KG and a heterogeneous news-graph learner with refinement and modality-attentive pooling for multimodal fake news detection.
Problem
Multimodal fake news detection is hard because static knowledge-graph entity embeddings inject noisy, content-irrelevant knowledge and fail to model correlations between multimodal semantics and knowledge entities.
Approach
Generate a compact, news-specific knowledge subgraph via an actor-critic RL agent walking the KG (rewarded by news relevance and detection correctness), then build a heterogeneous news graph with sentence, visual, and knowledge nodes, refine its topology with metric-learning enrichment and meta-path pruning, and pool hierarchically with modality attention.
Key results
- AKA-Fake reaches 0.919 accuracy on PolitiFact, 0.856 on GossipCop, and 0.858 on Pheme, improving accuracy by +4.7%, +2.4%, and +2.8% over the best baselines.
- Removing the KG input (w/o KG) drops PolitiFact accuracy from 0.919 to 0.869, and removing all knowledge and images drops it to 0.848.
- Replacing adaptive subgraph generation with static 2-order neighbors lowers PolitiFact accuracy to 0.889.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image and knowledge graph inputs (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.848 -> 0.919; 0.806 -> 0.856; 0.802 -> 0.858 | Helped |
| add image input (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.909 -> 0.919; 0.844 -> 0.856; 0.844 -> 0.858 | Helped |
| add knowledge graph input (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.869 -> 0.919; 0.819 -> 0.856; 0.817 -> 0.858 | Helped |
| use adaptive knowledge generation instead of static 1-order neighbors (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.879 -> 0.919; 0.818 -> 0.856; 0.826 -> 0.858 | Helped |
| use adaptive knowledge generation instead of static 2-order neighbors (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.889 -> 0.919; 0.825 -> 0.856; 0.830 -> 0.858 | Helped |
| use adaptive knowledge generation instead of top-10 semantic-similarity neighbors (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.879 -> 0.919; 0.817 -> 0.856; 0.828 -> 0.858 | Helped |
| add graph enrichment (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.879 -> 0.919; 0.811 -> 0.856; 0.815 -> 0.858 | Helped |
| add graph pruning (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.869 -> 0.919; 0.839 -> 0.856; 0.835 -> 0.858 | Helped |
| add graph refinement (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.869 -> 0.919; 0.823 -> 0.856; 0.831 -> 0.858 | Helped |
| use hierarchical pooling instead of mean pooling (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.909 -> 0.919; 0.843 -> 0.856; 0.839 -> 0.858 | Helped |
| use hierarchical pooling instead of max pooling (ablation) | AKA-Fake · PolitiFact, GossipCop, Pheme | accuracy: 0.899 -> 0.919; 0.845 -> 0.856; 0.849 -> 0.858 | Mixed |
Surprising findings
- Pheme results show KAHAN and CompNet-style baselines are close to AKA-Fake, and CAFE has higher F1 on Pheme (0.865) than some knowledge-enhanced baselines.
Limitations
- Small PolitiFact dataset (495 news) limits the strength of conclusions.
- Relies on external entity linking tools (TAGME, Faster R-CNN, visual entity linkage) whose errors propagate.
- Hyperparameters such as lambda and K show sensitivity requiring careful tuning.
Experimental setup
Datasets: PolitiFact GossipCop Pheme
Benchmarks: PolitiFact GossipCop Pheme
Metrics: Accuracy Precision Recall F1
Baselines: VGG-19 TextCNN BERT EANN MVAE SAFE SpotFake CAFE KAN CompNet KAHAN
Builds on: CompNet KAHAN CAFE TAGME CLIP
Closest papers by meaning
- External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.91)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.90)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.87)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.87)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.87)
- RaCMC: Residual-Aware Compensation Network with Multi-Granularity Constraints for Fake News Detection (AAAI 2025 · similarity 0.86)
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.