Research map / Graph learning / Graph self-supervised learning
GIN-SD: Source Detection in Graphs with Incomplete Nodes via Positional Encoding and Attentive Fusion
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 GIN-SD, a graph attention and Laplacian-positional-encoding framework with class-balanced loss for rumor source detection when a fraction of user data is missing.
Problem
Rumor source detection in social graphs assumes complete user data (states, forwarding frequency, reception times); real snapshots have incomplete nodes, which degrades existing detectors and causes source/non-source class imbalance.
Approach
Embeds Laplacian positional encodings of the infected subgraph plus user state and timestamp features to flag incomplete nodes, applies a GAT-style attentive fusion module to weight nodes by transmission capacity, and uses a class-balanced cross-entropy loss weighting source and non-source nodes by a constant ratio.
Key results
- With 10% incomplete nodes, GIN-SD reaches Acc 0.956 and F-Score 0.839 on Football, versus 0.897/0.729 for the best baseline IVGD.
- With 20% incomplete nodes, GIN-SD reaches Acc 0.921 and F-Score 0.694 on Enron across datasets, outperforming IVGD and SL-VAE on all eight graphs.
- Ablations show removing the class-balancing term (w/o B) drops Facebook F-Score from 0.761 to 0.223.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Laplacian positional encodings instead of zero-vector positional features (ablation) | GIN-SD · Facebook, 10% incomplete nodes | F-Score: 0.413 -> 0.761 | Helped |
| evaluate incomplete-source recognition using the P' criterion (ablation) | GIN-SD · Facebook, 10% incomplete nodes | reported accuracy: 0.426 | No clear effect |
| add the attention mechanism rather than uniform neighbor weighting (ablation) | GIN-SD · Facebook, 10% incomplete nodes | F-Score: 0.743 -> 0.761 | Helped |
| use learned attention rather than weighting smaller-degree nodes more (ablation) | GIN-SD · Facebook, 10% incomplete nodes | F-Score: 0.223 -> 0.761 | Helped |
| use learned attention rather than weighting larger-degree nodes more (ablation) | GIN-SD · Facebook, 10% incomplete nodes | F-Score: 0.716 -> 0.761 | Helped |
| add class balancing instead of setting the class weight to 1 (ablation) | GIN-SD · Facebook, 10% incomplete nodes | F-Score: 0.223 -> 0.761 | Helped |
Surprising findings
- Assigning higher attention to large-degree nodes (w/ AL) performs worse than uniform attention (w/o A) on the ablation variants, suggesting degree alone is not a good proxy for transmission capacity.
Limitations
- Attention heads and hidden sizes are reduced for medium and large graphs due to computational constraints.
- Evaluated only on simulated IC propagation with uniform forwarding probabilities, and the incomplete-node model is a simple random removal.
Experimental setup
Datasets: Football Jazz Facebook Twitch-ES LastFM Enron Github DBLP
Benchmarks: Football Jazz Facebook Twitch-ES LastFM Enron Github DBLP
Metrics: Accuracy F-Score
Baselines: LPSI EPA GCNSI SIGN GCSSI ResGCN IVGD SL-VAE
Builds on: Graph Attention Networks Laplacian positional encodings IVGD SL-VAE
Closest papers by meaning
- Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection (AAAI 2024 · similarity 0.85)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.84)
- Enhancing Robustness of Graph Neural Networks on Social Media with Explainable Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.83)
- Self-Discriminative Modeling for Anomalous Graph Detection (ICML 2025 · similarity 0.82)
- Backdoor Attack on Propagation-based Rumor Detectors (AAAI 2025 · similarity 0.82)
- Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.82)
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.