Research map / Graph learning / Graph self-supervised learning
Multi-Chain Graphs of Graphs: A New Approach to Analyzing Blockchain Datasets
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.
Introduces cross-chain Graphs-of-Graphs datasets of ERC20/BEP20 token transaction graphs (Ethereum, Polygon, BSC) with token-category labels and Jaccard-weighted inter-token global graphs, and benchmarks GNN and GoG models on classification, anomaly detection, and edge prediction.
Problem
Blockchain graph learning lacks large-scale, cross-chain datasets with hierarchical graph-level labels, and it is unclear whether Graphs-of-Graphs models perform well on large real-life transaction networks.
Approach
Model each token's transactions as a local graph and inter-token user overlap (Jaccard coefficient of shared addresses) as a global graph, forming a graphs-of-graphs dataset across three EVM chains with explorer-derived token category labels.
Key results
- SEAL achieves the best 3-class F1-macro on BSC (63.77) and up to 44% improvement over average non-GoG models in 5-class on Ethereum.
- On BSC anomaly detection, AnomalyDAE shows up to 3.34% AUC and 54.20% AP improvement over average structure-based methods, while on Ethereum and Polygon structure-based methods are superior.
- GoG models do not consistently outperform GNNs in global edge prediction; GoGNN reaches 66.94% accuracy on Ethereum, while performance on Polygon remains low across all models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GoG models instead of individual-graph GNNs for token classification | Token graph classification · Top-category classification across Ethereum, Polygon, and BSC | F1-macro: up to 28% and 16% improvements respectively in the 3-class category on BSC; up to 44% and 11% improvements respectively | Helped |
| Increase classification from 3 classes to 5 classes | Graph classification models · Ethereum, Polygon, and BSC token categories | F1-macro | Hurt |
| Use GoG models rather than individual-graph GNNs | Token graph classification · 5-class classification, especially minor classes | F1-macro | Mixed |
| Use graph anomaly detection methods instead of methods based on graph structural data | Token fraud detection · BSC | AUC: 3.34% | Helped |
| Use graph anomaly detection methods instead of methods based on graph structural data | Token fraud detection · Ethereum and Polygon | AUC | Hurt |
| Use GoG models instead of individual-graph GNNs | Global token link prediction · Ethereum, Polygon, and BSC | AUC | Mixed |
| Represent token graphs using DeepWalk (ablation) | GoG models and multivariate outlier detection methods · Token fraud detection | AUC | Mixed |
Surprising findings
- Graph outlier detection methods outperform structure-based methods on BSC but not on Ethereum and Polygon.
- Polygon remains hard for all model families, and temporal split results deteriorate sharply for Polygon.
Limitations
- Multiple accounts controlled by one entity can distort interaction patterns and connectivity metrics.
- Labels depend on blockchain explorer tags, which may be incomplete or misclassified, especially for fraud tokens.
- Existing GoG models underperform on minor-class classification and on Polygon, and node-degree features are weak for global edge prediction.
Experimental setup
Datasets: Ethereum ERC20 token transactions (Etherscan) Polygon token transactions (Polygonscan) BSC BEP20 token transactions (Bscscan)
Benchmarks: multi-class token classification (3-class, 5-class) graph anomaly detection (fraud token detection) global edge prediction
Metrics: F1-macro F1-micro AUC AP accuracy
Baselines: GCN GAT GIN ResidualGCN GraphSage SEAL GoGNN DVGGA COPOD IForest DIF VAE
Builds on: SEAL GoGNN DVGGA PyGOD PyOD
Closest papers by meaning
- Live Graph Lab: Towards Open, Dynamic and Real Transaction Graphs with NFT (NeurIPS 2023 · similarity 0.89)
- The Temporal Graph of Bitcoin Transactions (NeurIPS 2025 · similarity 0.89)
- EX-Graph: A Pioneering Dataset Bridging Ethereum and X (ICLR 2024 · similarity 0.88)
- BlockScan: Detecting Anomalies in Blockchain Transactions (NeurIPS 2025 · similarity 0.83)
- GraphLand: Evaluating Graph Machine Learning Models on Diverse Industrial Data (NeurIPS 2025 · similarity 0.81)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.81)
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.