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 advanced cryptocurrency fraud detection.

Unlike traditional blockchain datasets that rely strictly on transactional heuristics, this dataset is constructed using a Multimodal Data Fusion approach. It fuses On-Chain transactional behavior with off-chain contextual signals from three external modalities: Market Data, Reddit Sentiment, and Twitter/X Metrics.

The dataset consists of 35,272 unique Ethereum Externally Owned Accounts (EOAs). It exhibits a real-world class imbalance, with approximately 16.3% of the addresses labeled as fraudulent (scams/phishing) and the remaining 83.7% representing normal behavioral addresses.

Supported Tasks and Leaderboards

• tabular-classification: The dataset is intended to train models to classify Ethereum addresses as either Normal (0) or Scam (1).
• Fraud Detection Research: Evaluates the efficacy of multimodal data fusion versus traditional unimodal (on-chain only) anomaly detection.

Languages

Data features are numerical representations of behavior, market flow, and text sentiment. The underlying social data (Reddit/Twitter) from which sentiment features were derived was in English (en).

---

Dataset Structure

Data Instances

Each instance represents a singular, fused profile for a distinct Ethereum wallet address (address_level_fused.parquet). The features represent aggregated temporal behaviors.

Data Fields

The dataset contains a total of 31 specific features alongside the address identifier and the target label.

Identifiers & Labels:

• address (string): The 42-character Ethereum Externally Owned Account address.
• is_scam (int): The binary target variable (1 = Scam, 0 = Normal).

On-Chain Features (14): Transactional heuristics including flows, burst activity, and wallet lifecycle.

• Examples: eth_recv_mean, eth_net_flow_max, burst_max_tx_5m_mean, active_days.

Market Features (5): Macroeconomic indicators at the exact time of the address's active transactions.

• Examples: market_daily_return_mean, market_intraday_volatility_mean.

Reddit Features (3): Community sentiment and scam mentions correlated temporally with the address's activity windows.

• Examples: reddit_total_fraud_mentions_mean, reddit_sentiment_std_mean.

Twitter/X Features (9): Social velocity and sentiment surrounding the address or related smart contracts during its active periods.

• Examples: twitter_avg_retweets_mean, twitter_avg_positive_mean.

(For a full schema mapping, refer to the address_features_metadata.json located in the accompanying model repository).

Data Splits

The dataset (address_level_fused.parquet) contains all 35,272 records. In the FuseChain study, this was split using an 80/20 stratified sampling method to preserve the ~16.3% minority class prevalence:

• Training Set: 28,217 profiles
• Test Set: 7,055 profiles

---

Dataset Creation

Curation Rationale

Detecting modern Web3 fraud requires more than just analyzing token flows. Scammers manipulate market conditions and leverage social media (X/Reddit) to execute phishing campaigns and rug pulls. This dataset was created to prove that fusing these OSINT (Open Source Intelligence) modalities significantly improves recall and precision in fraud detection.

Source Data

This fused dataset builds upon raw datasets sourced from the community and external APIs. The raw data foundations include:

• On-chain Data: Derived from the base transaction structures available in Ethereum Fraud Dataset by Activity (Fesevu).
• Market Data: Historical Ether (ETH) pricing and trading volumes downloaded as CSVs from CoinGecko.
• Reddit Data: Community sentiment signals derived from The Reddit Ethereum Dataset (pavellexyr) on Kaggle.
• Twitter/X Data: Social engagement and sentiment metrics derived from the Cryptocurrency Tweets with Sentiment Analysis (fabioturazzi) dataset on Kaggle.

Preprocessing and Feature Engineering

1. Temporal Alignment: Daily transaction summaries were mapped to the exact daily closing market prices and daily aggregated social sentiment from the raw sources.
2. Address-Level Aggregation: Daily behavior was aggregated into singular "profiles" using statistical functions (Mean, Max, Total sum).
3. Feature Reduction: The initial feature space of >60 features was condensed to the final 31 highly discriminative features using a sequential pipeline including Variance Thresholds, Pairwise Spearman Correlation filters, and Mutual Information ranking.

---

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to the ongoing effort to secure Decentralized Finance (DeFi) ecosystems. By proving the value of off-chain signals, it provides a foundation for more robust Web3 security tools capable of identifying malicious actors before extensive financial damage occurs.

Known Limitations

• Temporal Decay: Scam behavioral patterns evolve rapidly. While the historical data provides a solid foundation, real-time applicability requires continuous retraining.
• Source Dependencies: The features are inherently tied to the structures of the underlying raw Kaggle/HuggingFace datasets from which they were fused.

---

Citation Information

If you use this dataset in your research, please attribute the FuseChain project.

@misc{fusechain2025,
  title={FuseChain: Multimodal Ethereum Fraud Detection using On-chain and Off-chain Signals},
  author={[Nileshka Fernando]},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Nileshka/fusechain-data}}
}