Research map / Graph learning / Knowledge and temporal graphs
Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference
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 HML, a fake news detector that infers latent news cascades via an event-based Hawkes process and builds a multi-edge heterogeneous news graph, combined with self-supervised multimodal contrastive learning and a graph transformer, to detect complex multimodal fake news without retweet structure.
Problem
Fake news detection on short-video platforms (TikTok, Instagram, YouTube) where real propagation/social cascades are inaccessible, and where multimodal features are simply spliced, losing inter-modal relationships.
Approach
Infers latent social influence among news items of the same event using an improved Hawkes process (with a diffusion-style noising and Minkowski-distance similarity for cross-event influence), builds a heterogeneous news graph from attribute similarities (no retweets), and learns multimodal features via unimodal augmentation and cross-modal contrastive losses before a graph transformer classifier.
Key results
- HML reaches 89.14% accuracy and 89.22% F1 on FakeSV, versus 85.35% accuracy for the prior best FakingRecipe.
- On FVC, HML reports 91.02% accuracy and 90.58% F1, versus 89.28% accuracy for MMVD.
- As a plug-in network inference module, it improves EANN on Twitter accuracy from 64.86% to 74.72%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HML instead of FakingRecipe | FakeSV fake news detection · FakeSV | accuracy: +3.79% | Helped |
| use HML instead of FakingRecipe | FakeSV fake news detection · FakeSV | F1: +4.39% | Helped |
| use HML instead of MMVD | FVC fake news detection · FVC | accuracy: +1.74% | Helped |
| use HML instead of MMVD | FVC fake news detection · FVC | F1: +0.12% | Helped |
| add Stage 1 latent network inference (ablation) | HML · FakeSV | accuracy: 81.68 -> 89.14 | Helped |
| add event-based latent inference (ablation) | HML · FakeSV | accuracy: 86.84 -> 89.14 | Helped |
| add heterogeneous edge-type construction (ablation) | HML · FakeSV | accuracy: 88.70 -> 89.14 | Helped |
| add Stage 2 multimodal content learning (ablation) | HML · FakeSV | accuracy: 78.57 -> 89.14 | Helped |
| add unimodal content augmentation (ablation) | HML · FakeSV | accuracy: 88.80 -> 89.14 | Helped |
| add cross-modal contrast enhancement (ablation) | HML · FakeSV | accuracy: 88.40 -> 89.14 | Helped |
| add network inference | EANN · Twitter and Weibo | accuracy: +9.86%; +5.33% | Helped |
| add network inference | MCAN · Twitter and Weibo | accuracy: +2.90%; +0.31% | Helped |
Surprising findings
- Prompting LLMs (Doubao, GPT-4o) performs poorly on multimodal fake news detection, with GPT-4o at 50.72% accuracy on FVC.
Limitations
- Relies on the assumption of no prior knowledge of the real network, using a hand-set similarity threshold to binarize edges.
- LLM prompting baselines are shown to be weak, and the paper does not adapt LLMs for this task.
Experimental setup
Datasets: FakeSV FVC Twitter Weibo
Benchmarks: FakeSV FVC Twitter Weibo
Metrics: Accuracy Precision Recall F1
Baselines: BERT VGGish VGG19 C3D Serrano et al. FANVM SV-FEND MMVD NEED FakingRecipe Doubao GPT-4o
Builds on: Hawkes process graph transformer focal loss EANN
Closest papers by meaning
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.90)
- External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.89)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.88)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.86)
- GAMC: An Unsupervised Method for Fake News Detection Using Graph Autoencoder with Masking (AAAI 2024 · similarity 0.86)
- Unveiling Implicit Deceptive Patterns in Multi-Modal Fake News via Neuro-Symbolic Reasoning (AAAI 2024 · similarity 0.85)
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.