Research map / Graph learning / Graph self-supervised learning
Graph is a Substrate Across Data Modalities
Ziming Li, Xiao-Ming Wu, Zehong Wang, Jiazheng Li, Yijun Tian, Jinhe Bi, Yunpu Ma, Yanfang Ye, Chuxu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add interleaved role-based training on top of unified multi-task schema (ablation) | Unified multi-task (Qwen3-VL-2B) · Shortest path | accuracy: 45.02 -> 48.59 | Improved |
| add interleaved role-based training on top of unified multi-task schema (ablation) | Unified multi-task · HiEvent | F1: 19.37 -> 25.15 | Improved |
| use unified schema instead of native formats in task-isolated training (ablation) | single-task training · Shortest path / bipartite matching | accuracy: SP 38.27 -> 37.14; BM 92.05 -> 85.98 | Hurt |
| use unified schema instead of native formats in multi-task training (ablation) | multi-task training · Shortest path | accuracy: 41.27 -> 45.02 | Improved |
| add interleaving without unified schema (ablation) | Naive multi-task · MAVEN-Causal / Cycle detection | F1 / accuracy: SP 41.27 -> 43.83; CD 94.72 -> 93.86 | Mixed |
| use XML-style serialization instead of the proposed schema realization (ablation) | G-Substrate · MAVEN-Temporal etc. | F1: 42.68 -> 36.40 | Hurt |
| use natural-language serialization instead of proposed schema (ablation) | G-Substrate · Shortest path | accuracy: 48.59 -> 44.80 | Hurt |
| pretrain on event graphs only (unified schema) before scene graph generation (ablation) | Qwen3-VL-2B · Scene graph generation PCIs | R@50: 19.10 -> 21.47 | Improved |
| replace reused graphs with structurally incorrect variants (ablation) | interleaved training · GAR/MGD/SGG/ERE | delta vs unified multi-task: gains of +0.62 to +2.23 become -1.20 to -2.03 in Figure 5 | Hurt |
| raise interleaved data ratio from 50% to 100% (ablation) | G-Substrate · GAR | accuracy: 94.5 -> 93.4 | Hurt |
| add 50% interleaved data instead of none (ablation) | G-Substrate · ERE | averaged F1: 38.0 -> 40.2 | Improved |
| corrupt 20% of reused graph edges (ablation) | G-Substrate · MGD | BLEU-4: 50.74 vs clean NMT 48.11 | Improved |
Closest papers by meaning
- Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.89)
- ICLR: In-Context Learning of Representations (ICLR 2025 · similarity 0.89)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.88)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.87)
- Microstructures and Accuracy of Graph Recall by Large Language Models (NeurIPS 2024 · similarity 0.87)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.87)
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.
