Nileshka/fusechain-data

Dataset Card for FuseChain Multimodal Ethereum Fraud Dataset

Dataset Summary

The FuseChain Multimodal Ethereum Fraud Dataset is a novel, address-level dataset designed for Ethereum fraud detection research. Unlike traditional blockchain datasets that rely solely on transactional heuristics, this dataset is constructed using a multimodal data fusion approach that integrates on-chain transactional behaviour with off-chain contextual signals from three external modalities: market data, Reddit community signals, and Twitter sentiment metrics.

The dataset consists of 35,272 unique Ethereum Externally Owned Accounts (EOAs), reflecting a real-world class distribution of approximately 16.3% fraudulent addresses and 83.7% normal addresses.

---

Supported Tasks

• Tabular Classification: Train models to classify Ethereum addresses as either Normal (0) or Scam (1)
• Fraud Detection Research: Evaluate the efficacy of multimodal data fusion versus traditional on-chain only approaches
• Modality Contribution Analysis: Assess the independent and combined contribution of each off-chain data source through ablation studies

---

Languages

Features are numerical representations of behaviour, market conditions, and text sentiment. The underlying social media data from which sentiment features were derived was in English.

---

Dataset Structure

Data Instances

Each instance represents a single fused behavioural profile for a distinct Ethereum EOA address stored in address_level_fused.parquet. Features represent aggregated temporal behaviours across the address's entire active period.

Data Fields

The dataset contains 31 features alongside the address identifier and target label, spanning four modalities.

Identifiers and Labels:

• address (string): The 42-character Ethereum EOA address
• is_scam (int): Binary target variable — 1 = Scam, 0 = Normal

On-Chain Features (14): Transactional behavioural heuristics including ETH flows, burst activity, session patterns, and wallet lifecycle metrics.

"active_days", "normal_sent_cnt_mean", "normal_sent_cnt_max", "normal_recv_cnt_mean", "normal_recv_cnt_max", "normal_total_cnt_mean", "normal_total_cnt_max", "eth_sent_sum_mean", "eth_net_flow_mean", "eth_net_flow_max", "sessions_cnt_mean", "burst_max_tx_5m_mean", "burst_max_tx_5m_max", "days_since_last_activity_mean"

Market Features (5): Macroeconomic indicators from CoinMarketCap capturing daily ETH price dynamics and trading conditions during the address's active period.

"market_daily_return_mean", "market_volatility_7d_mean", "market_volume_spike_mean", "market_momentum_7d_mean", "market_intraday_volatility_mean"

Reddit Features (3): Community discussion signals and fraud-related discourse metrics from Ethereum-related Reddit communities, temporally aligned to the address's active period.

"reddit_fraud_mention_ratio_mean", "reddit_total_fraud_mentions_mean", "reddit_avg_sentiment_mean"

Twitter Features (9): Social engagement and sentiment metrics from Ethereum-related Twitter activity during the address's active period.

"twitter_total_activity_mean", "twitter_avg_sentiment_mean", "twitter_avg_negative_mean", "twitter_avg_positive_mean", "twitter_negative_ratio_mean", "twitter_fraud_mention_ratio_mean", "twitter_avg_likes_mean", "twitter_avg_retweets_mean", "twitter_avg_replies_mean"

For the full feature schema, refer to address_features_metadata.json in the accompanying model repository.

---

Data Splits

The full dataset of 35,272 records was split using 80/20 stratified sampling to preserve the 16.3% minority class distribution:

Split	Addresses	Normal	Scam
Full Dataset	35,272	29,530 (83.7%)	5,742 (16.3%)
Training Set	28,217	23,623	4,594
Test Set	7,055	5,907	1,148

---

Dataset Creation

Curation Rationale

Modern Ethereum fraud detection requires more than analysing token flows. Scammers exploit social media platforms and market conditions to coordinate phishing campaigns, rug pulls, and pump-and-dump schemes. This dataset was created to demonstrate that fusing on-chain transaction features with off-chain contextual signals significantly improves fraud detection performance, providing a reusable research artifact for the wider blockchain security community.

Source Data

This fused dataset is built upon raw data sourced from the following publicly available datasets:

• On-Chain Data: Ethereum Fraud Dataset by Activity (Ishmukhamedov, 2025) hosted on Hugging Face
• Market Data: Historical ETH OHLCV pricing and trading volume data from CoinMarketCap
• Reddit Data: The Reddit Ethereum Dataset (pavellexyr) on Kaggle
• Twitter Data: Cryptocurrency Tweets with Sentiment Analysis (fabioturazzi) on Kaggle

Analysis Window

All datasets were temporally aligned within the shared research window of 2017-01-01 to 2021-02-10, constrained by the coverage of the Twitter dataset.

Preprocessing and Feature Engineering

• EOA Filtering: Dataset restricted to Externally Owned Accounts, excluding smart contract addresses
• Zero-Activity Removal: Days with no meaningful transaction activity were removed
• Temporal Alignment: All data sources aligned at daily granularity using a shared date key
• Feature Engineering: Five market features, four Reddit features, and eleven Twitter features derived from daily aggregated signals
• Address-Level Aggregation: Daily records aggregated into static address profiles using mean, maximum, and sum statistics for on-chain features, and mean only for off-chain features
• Feature Reduction: Initial feature space reduced to 31 final features using two-stage Spearman rank correlation filtering applied at both the daily and address levels

---

Considerations for Using the Data

Social Impact

This dataset contributes to securing Decentralised Finance ecosystems by demonstrating the value of off-chain contextual signals for fraud detection. It provides a foundation for future research into more robust Web3 security tools capable of identifying malicious addresses before significant financial damage occurs.

Known Limitations

• Temporal Coverage: The dataset covers 2017-2021 and may not reflect contemporary fraud patterns such as NFT-related scams or cross-chain bridge exploits
• Global Social Signals: Reddit and Twitter features represent global daily Ethereum community activity rather than signals directly attributable to individual addresses
• Label Noise: Address labels are sourced from community-reported databases which may contain mislabelled or unreported addresses
• Static Dataset: The dataset represents a fixed historical snapshot and requires retraining on updated data for contemporary applicability

---

Citation

If you use this dataset in your research, please cite the FuseChain project:

@misc{fusechain2026,
  title={FuseChain: Ethereum Fraud Detection via Multimodal Signal Fusion},
  author={Fernando, Nileshka},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Nileshka/fusechain-data}}
}

---

Related Resources

• Code Repository & Pipeline: FuseChain GitHub
• Model: Model Repository