Research map / Graph learning / Graph self-supervised learning
IOHunter: Graph Foundation Model to Uncover Online Information Operations
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 IOHunter, a graph foundation model that fuses frozen multilingual SBert text embeddings with degree-based graph embeddings via cross-attention before a GCN/SAGE classifier, to detect IO drivers on Twitter across six country datasets.
Problem
Detecting users who drive state-sponsored information operations on social media, including generalization to new IOs (cross-IO/cross-country) and low-label regimes.
Approach
Combine a frozen multilingual sentence-embedding model (SBert) for user content with degree-bucket one-hot graph features from a fused similarity network, blend them with a cross-attention gate, and train a GNN classifier; pretraining across countries enables transfer to unseen IOs.
Key results
- IOHunter averages 96.43 Macro-F1 across six countries versus 87.77 for node2vec+RF and 88.80 for the concatenation GNN.
- Pretraining on other countries and fine-tuning on 0.1% of the test country improves Macro-F1 by 10.25% on average over training on 0.1% alone.
- Gains over node2vec+RF range from +1.8% (UAE) to +20% (Iran) in supervised setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IOHunter instead of the strongest listed baseline | IO detection · UAE, Cuba, Russia, Venezuela, Iran, and China | Macro-F1: 96.43 vs 87.77 average | Helped |
| add cross-attention to fuse graph and text features instead of simple concatenation (ablation) | IOHunter · UAE, Cuba, Russia, Venezuela, Iran, and China | Macro-F1: 88.80 -> 96.43 average | Helped |
| add the graph modality (ablation) | IOHunter · UAE, Cuba, Russia, Venezuela, Iran, and China | Macro-F1: 86.76 -> 96.43 average | Helped |
| add the text modality (ablation) | IOHunter · UAE, Cuba, Russia, Venezuela, Iran, and China | Macro-F1: 88.53 -> 96.43 average | Helped |
| pretrain on other countries before target-country evaluation | IOHunter · Cross-country detection | Macro-F1: +5.69% | Mixed |
| fine-tune the pretrained model on 0.1% of target-country training data | IOHunter · Cross-country detection | Macro-F1: +10.25% | Helped |
Surprising findings
- Graph-only and text-only variants each excel on different countries, with the text-only variant far worse on UAE than graph-only, suggesting modality strength varies by IO.
- Simple concatenation of graph and text features yields high variance across seeds on some countries (e.g., Cuba ±17.9).
Limitations
- Evaluated only on Twitter IO datasets from six countries; generalization to other platforms is untested.
- Cross-IO results show high variance and gains depend on the country; performance on some countries (China) remains modest.
- Relies on a fixed similarity-network construction and frozen SBert text encoder.
Experimental setup
Datasets: Twitter IO datasets from prior work covering UAE, Cuba, Russia, Venezuela, Iran, China
Benchmarks: Twitter IO driver datasets (UAE, Cuba, Russia, Venezuela, Iran, China)
Metrics: Macro-F1
Baselines: NodePruning node2vec+RF SBert+MLP GNN (GCN/Sage concatenation)
Builds on: node2vec GCN GraphSAGE SBert PyTorch Geometric
Closest papers by meaning
- Adversarial Socialbots Modeling Based on Structural Information Principles (AAAI 2024 · similarity 0.82)
- SNS-Bench: Defining, Building, and Assessing Capabilities of Large Language Models in Social Networking Services (ICML 2025 · similarity 0.81)
- Tracking and Identifying International Propaganda and Influence Networks Online (AAAI 2025 · similarity 0.81)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.81)
- Intruding with Words: Towards Understanding Graph Injection Attacks at the Text Level (NeurIPS 2024 · similarity 0.81)
- MSR: A Multifaceted Self-Retrieval Framework for Microscopic Cascade Prediction (AAAI 2025 · similarity 0.80)
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.