Research map / Graph learning / Graph self-supervised learning
EX-Graph: A Pioneering Dataset Bridging Ethereum and X
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 EX-Graph, a dataset linking Ethereum transaction graphs with X (Twitter) follower graphs via 30,667 verified OpenSea matching links, and shows X features improve Ethereum link prediction, wash-trading detection, and matching link prediction.
Problem
Blockchain datasets are exclusively on-chain, and Ethereum addresses have few informative node features; prior on-/off-chain matching datasets are small or unreliable (ENS-based) or lack the on-chain graph.
Approach
Bridge Ethereum transaction graph and X follower graph using verified Ethereum-X account links from OpenSea profiles, then use X profile BERT embeddings and follower counts as off-chain node features to enhance on-chain tasks.
Key results
- Adding X features raised AUC-ROC for Ethereum link prediction across models, e.g., APPNP about 7% and DAGNN 8% increase in AUC-ROC (up to 8% peak per abstract).
- For wash-trading detection, X features boosted recall by 18% in GraphSage and 13% in GAT.
- Matching link prediction reached up to 0.74 AUC-ROC with DAGNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add X account features | Ethereum link prediction with APPNP · Ethereum link prediction | AUC-ROC: 0.82±0.01 -> 0.89±0.02 | Helped |
| Add X account features | Ethereum link prediction with DAGNN · Ethereum link prediction | AUC-ROC: 0.79±0.00 -> 0.87±0.00 | Helped |
| Add X account features | Ethereum link prediction with DAGNN · Ethereum link prediction | Recall: 0.67±0.01 -> 0.77±0.01 | Helped |
| Add X account features | Ethereum link prediction with TAGCN · Ethereum link prediction | AUC-ROC: 0.81±0.01 -> 0.88±0.00 | Helped |
| Add X account features | Wash-trading detection with GraphSage · Wash-trading detection | Recall: 0.52±0.01 -> 0.70±0.02 | Helped |
| Add X account features | Wash-trading detection with GAT · Wash-trading detection | Recall: 0.59±0.00 -> 0.72±0.00 | Helped |
| Add X account features | Wash-trading detection with GAT · Wash-trading detection | Precision: 0.80±0.01 -> 0.74±0.01 | Hurt |
| Add X account features | Wash-trading detection with GCN · Wash-trading detection | AUC-ROC: 0.79±0.01 -> 0.79±0.01 | No clear effect |
Surprising findings
- Only three of the 1,445 wash-trading addresses match to X accounts, reflecting how hidden wash-trading activity is.
- Some models (e.g., GAT, GATv2) show lower precision with X features on wash-trading detection despite higher recall.
Limitations
- Few X-matched Ethereum addresses (30,667) and very few matched wash-trading addresses (3), limiting label coverage.
- Severe class imbalance for wash-trading addresses; balanced sampling used for evaluation.
- X data collection limited by API rate limits and anonymization of handles; BERT features only from profile text.
Experimental setup
Datasets: EX-Graph Ethereum transaction graph (NFT transactions, Mar-Aug 2022) EX-Graph X following graph OpenSea verified Ethereum-X matching links Dune wash-trading labels (1,445 addresses)
Benchmarks: EX-Graph Ethereum link prediction EX-Graph wash-trading addresses detection EX-Graph matching link prediction
Metrics: AUC-ROC Precision Recall F1-score Accuracy
Baselines: DeepWalk Node2Vec GGNN GCN GAT GATv2 GraphSage APPNP TAGCN Cluster-GCN DAGNN HIN2Vec
Builds on: DeepWalk Node2Vec GraphSAGE BERT DGL
Closest papers by meaning
- Multi-Chain Graphs of Graphs: A New Approach to Analyzing Blockchain Datasets (NeurIPS 2024 · similarity 0.88)
- The Temporal Graph of Bitcoin Transactions (NeurIPS 2025 · similarity 0.86)
- BlockScan: Detecting Anomalies in Blockchain Transactions (NeurIPS 2025 · similarity 0.80)
- TEG-DB: A Comprehensive Dataset and Benchmark of Textual-Edge Graphs (NeurIPS 2024 · similarity 0.79)
- Approval-Based Committee Voting in Practice: A Case Study of (over-)Representation in the Polkadot Blockchain (AAAI 2024 · similarity 0.77)
- EVMbench: Evaluating AI Agents on Smart Contract Security (ICML 2026 · similarity 0.77)
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.