Research map / Graph learning / GNN architectures
Self-Explainable Graph Transformer for Link Sign Prediction
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 SE-SGformer, a self-explainable graph Transformer for link sign prediction that uses signed random walk encoding and explains predictions via K-nearest positive and K-farthest negative neighbors.
Problem
Signed GNNs for link sign prediction are not explainable, and post-hoc explanations may be biased; there is also a scarcity of negative neighbors in signed graphs.
Approach
A graph Transformer for signed graphs with centrality, adjacency, and signed random walk encodings, whose embeddings feed a distance-based decision process that predicts edge signs by comparing an edge's distance to a node's K-nearest positive and K-farthest negative neighbors, which also serve as explanations.
Key results
- SE-SGformer reaches 90.03% accuracy on Bitcoin-OTC and 89.88% on Bitcoin-Alpha, versus 88.22% and 87.96% for the best baselines (SGCN).
- On KuaiRec, SE-SGformer reaches 85.60% accuracy, while SGCN reaches 85.11%; on Epinions it is lower (72.84%) than SGCN (86.97%).
- Explanation precision@40 is 75.19% on Bitcoin-OTC, 94.47% on Bitcoin-Alpha and 76.07% on Amazon-music, above all listed baselines (at most 61.70%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · Amazon-music | accuracy: 78.26 ± 1.52 -> 79.20 ± 0.23 | Helped |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · Epinions | accuracy: 86.97 ± 1.53 -> 72.84 ± 1.78 | Hurt |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · KuaiRand | accuracy: 62.85 ± 0.05 -> 56.89 ± 0.12 | Hurt |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · KuaiRec | accuracy: 85.11 ± 0.11 -> 85.60 ± 0.05 | Helped |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · WikiRfa | accuracy: 78.69 ± 1.06 -> 79.99 ± 0.08 | Helped |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · WikiElec | accuracy: 79.63 ± 2.82 -> 80.63 ± 0.08 | Helped |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · Bitcoin-OTC | accuracy: 88.22 ± 0.69 -> 90.03 ± 0.35 | Helped |
| Use SE-SGformer instead of the strongest baseline | Link sign prediction · Bitcoin-Alpha | accuracy: 87.96 ± 0.38 -> 89.88 ± 0.40 | Helped |
| Use SE-SGformer instead of the strongest baseline | Neighbor explanations · Bitcoin-OTC | Precision@40: 57.25 ± 0.15 -> 75.19 ± 0.13 | Helped |
| Use SE-SGformer instead of the strongest baseline | Neighbor explanations · Bitcoin-Alpha | Precision@40: 55.43 ± 0.15 -> 94.47 ± 0.58 | Helped |
| Use SE-SGformer instead of the strongest baseline | Neighbor explanations · Amazon-music | Precision@40: 61.70 ± 0.08 -> 76.07 ± 0.20 | Helped |
| Add centrality encoding (ablation) | SE-SGformer · Bitcoin-OTC and Bitcoin-Alpha | accuracy | Helped |
Surprising findings
- SE-SGformer underperforms SGCN on Epinions (72.84% vs 86.97%) and KuaiRand (56.89% vs 62.85%).
- SIGformer, a signed graph Transformer, performs poorly on most datasets (e.g., 57.07% on Epinions).
Limitations
- Performance is below SGCN on several datasets such as Epinions and KuaiRand.
- Explanation ground truth is synthetic, derived from the model's own learned embeddings rather than human-verified rationales.
- Number of Transformer layers is fixed to one in the main setting; deeper models reduce performance.
Experimental setup
Datasets: Bitcoin-OTC Bitcoin-Alpha WikiElec WikiRfa Epinions KuaiRand KuaiRec Amazon-music
Benchmarks: Bitcoin-OTC Bitcoin-Alpha WikiElec WikiRfa Epinions KuaiRand KuaiRec Amazon-music
Metrics: accuracy precision@K
Baselines: GCN GAT SGCN SNEA SGCL SIGformer
Builds on: Graphormer SGCN SRWR SE-GNN SIGformer
Closest papers by meaning
- DropEdge not Foolproof: Effective Augmentation Method for Signed Graph Neural Networks (NeurIPS 2024 · similarity 0.91)
- Structure Balance and Gradient Matching-Based Signed Graph Condensation (AAAI 2025 · similarity 0.87)
- GOAt: Explaining Graph Neural Networks via Graph Output Attribution (ICLR 2024 · similarity 0.87)
- Sign is Not a Remedy: Multiset-to-Multiset Message Passing for Learning on Heterophilic Graphs (ICML 2024 · similarity 0.87)
- Bridging Theory and Practice in Link Representation with Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- A Signed Graph Approach to Understanding and Mitigating Oversmoothing (NeurIPS 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.