fesevu/ethereum_fraud_dataset_by_activity

🕸 Ethereum Address Behavior Dataset — GNN + LSTM (Fraud Detection)

This dataset is designed for fraud detection on Ethereum addresses using a dual-modality approach:

• Graph Neural Networks (GNN): transaction graph structure.
• Recurrent Models (LSTM/Transformers): time-series of address features.

The dataset is built from:

• Ethereum public BigQuery dataset (bigquery-public-data.crypto_ethereum.transactions).
• Etherscan labels + custom scam labels.
• Balanced address list of ~115k addresses (scam vs non-scam, contracts vs EOAs).

The dataset is also available on Kaggle: Ethereum Fraud Dataset by Activity

📦 Dataset Collection Pipeline

To reproduce or customize the dataset, use the instructions and code in the eth-fraud-dataset-pipeline repository.
That repository provides:

• Scripts for downloading raw data from public sources (BigQuery, Etherscan, curated scam lists).
• Code for merging, deduplicating, and balancing address labels.
• Tools for building the GNN and LSTM datasets (parquet files, mappings, targets).
• Utilities for generating checksums and manifests for data integrity.

You must run the provided scripts to generate the dataset locally; the data files are not stored in the GitHub repository.

---

📂 Repository Structure

final/ ├─ gnn_dataset/ # GNN dataset (edges, meta, labels, mapping, targets) │ ├─ edges_all/edges.parquet │ ├─ edges_by_week/week=YYYY-Www/edges.parquet │ ├─ edges_by_month/month=YYYY-MM/edges.parquet │ ├─ meta/{week,month}_window_meta.parquet │ ├─ labels/targets_global.parquet │ ├─ mapping/address_id_map_labels.parquet │ ├─ targets/{week,month}_targets.parquet │ └─ README.md └─ lstm_dataset/ # LSTM dataset (daily → weekly → monthly aggregations) ├─ daily_filtered.parquet ├─ weekly.parquet ├─ monthly.parquet └─ README.md

• gnn_dataset/ → GNN dataset (graph edges, slices, labels, mapping).
• lstm_dataset/ → LSTM dataset (tabular features, time-series).

---

🔑 Synchronization Between GNN and LSTM

• Both use the same address universe (node_id mapping).
• Both use the same time windows:
  • ISO weeks (YYYY-Www) from gnn_dataset/meta/week_window_meta.parquet.
  • Months (YYYY-MM) from gnn_dataset/meta/month_window_meta.parquet.

---

📂 Raw Data

Alongside the processed datasets, we also provide the raw parquet exports (all parquet files are compressed with Zstandard (zstd)):

final/ ├─ GNN/parquet/ # raw transaction parquet chunks for GNN │ ├─ transactions_daily_part-00000.parquet │ ├─ transactions_daily_part-00001.parquet │ └─ ... ├─ LSTM/parquet/ # raw daily features parquet for LSTM │ ├─ daily_final_part-00000.parquet │ ├─ daily_final_part-00001.parquet │ └─ ... ├─ addr_labels_balanced.csv # balanced address list with labels ├─ addr_labels_balanced.csv # balanced subset with labels (used in GNN + LSTM)

---

Contents

• GNN/parquet/ — raw transaction-level parquet files, containing:
  • from_address, to_address (STRING, lowercase hex)
  • block_number (INT64)
  • timestamp (TIMESTAMP, UTC)
  • value_wei, tx_fee_wei (NUMERIC in source, stored as string later)
  • nonce, input_data_size, contract_creation, tx_hash, day
• LSTM/parquet/ — raw daily activity parquet files (address-day features before filtering).
• addr_labels_big.csv — initial large list of Ethereum addresses (>1M), with scam/contract metadata, not used directly (later downsampled & balanced).
• addr_labels_balanced.csv — final balanced list of ~115k addresses (scam vs non-scam, contract vs EOA), used for both GNN and LSTM datasets.

All parquet files in this dataset are compressed using Zstandard (zstd) for efficient storage and fast access.

These files are the starting point for the preparation scripts:

• build_unified_dataset.py → creates gnn_dataset/ (GNN).
• build_lstm_dataset_lowmem.py → creates lstm_dataset/ (LSTM).

---

⚖️ Labels

• Source: Etherscan tags + curated scam lists.
• Balanced across:
  • Scam vs Non-Scam
  • Contract vs EOA
• Provided in:
  • gnn_dataset/labels/targets_global.parquet
  • gnn_dataset/mapping/address_id_map_labels.parquet

---

Address Label Files

Both addr_labels_big.csv (full set) and addr_labels_balanced.csv (balanced subset) share the same schema:

Field	Type	Units	Description
address	STRING	hex	Ethereum address (0x..., lowercase).
is_scam	INT64	0/1	Scam label: 1 = scam, 0 = non-scam.
description	STRING	—	Free-text description (e.g. "Verified", "Phishing").
activity_start_ts	TIMESTAMP	UTC	First observed activity timestamp.
activity_end_ts	TIMESTAMP	UTC	Last observed activity timestamp.
is_contract	INT64	0/1	Address type: 1 = smart contract, 0 = EOA.

• addr_labels_big.csv — ~1M+ raw addresses with scam/contract metadata, not used directly (later downsampled and balanced).
• addr_labels_balanced.csv — final balanced subset (~115k addresses, scam vs non-scam, contract vs EOA), used in both GNN and LSTM datasets.

---

📦 Use Cases

• Graph ML: Train static embeddings (GraphSAGE, Node2Vec) or temporal GNNs.
• Sequence ML: Train LSTM/Transformer on address time-series features.
• Fusion: Combine GNN embeddings and LSTM features via node_id.
• Fraud detection: Predict scam addresses, contracts vs EOAs.

---

🛠 Collection Details

• Source: Ethereum mainnet via BigQuery.
• Labels: from Etherscan + custom curated lists.
• Timezone: UTC.
• ETH amounts stored as Decimal(38,9), exported as strings for precision.
• Data preparation optimized for BigQuery + Polars, fits in 12–24 GB RAM.

🗂 Source Datasets

The address list and labels (scam/non-scam, description) were compiled from the following public datasets:

• Primary sources:

  • xblock.pro Dataset #13
  • xblock.pro Dataset #25
  • xblock.pro Dataset #50
  • PTXPhish
  • Phishing Contract Sigmetrics
  • Etherscan Open Source Contract Codes
  • MyEtherWallet Ethereum Lists
  • EtherScamDB
  • CryptoScamDB Blacklist
  • ScamSniffer Scam Database
  • Forta Network Labelled Datasets
  • Kaggle: Labelled Ethereum Addresses
  • Etherscan Labels
  • Kaggle: Ethereum Fraud Detection Dataset
  • Ethereum Fraud Datasets
  • Kaggle: Ponzi Scheme Contracts
  • Ethereum Fraud Detection

• Integration:

  • Addresses and labels from these sources were merged and deduplicated.
  • The final balanced address list (~115k addresses) was constructed based on these datasets.