Kshitijbhatt1998/ieee-fraud-detection-pipeline-features

IEEE-CIS Fraud Detection — Model-Ready Feature Dataset

Dataset Summary

This dataset contains 590,540 financial transactions from the IEEE-CIS Fraud Detection competition, processed through a production-grade data pipeline into a clean, labeled, model-ready feature table.

The pipeline adds 16 engineered features on top of the original IEEE-CIS columns — including card-level velocity, email domain risk rates, and amount anomaly ratios — ready for direct use in fraud detection model training.

This is the output of the pipeline described in fintech-fraud-pipeline.

Dataset Details

Property	Value
Rows	590,540
Features	58
Label column	is_fraud (binary: 0 = legit, 1 = fraud)
Fraud rate	~3.5%
Date range	~6 months (Nov 2017 – May 2018)
Delivery format	Parquet

Supported Tasks

Binary classification — predicting whether a financial transaction is fraudulent.

Suitable for training and benchmarking: gradient boosting models (XGBoost, LightGBM), neural networks, or anomaly detection systems.

Data Source

Original data: IEEE-CIS Fraud Detection (Kaggle, 2019).
All features are anonymized per the original dataset terms. Transaction patterns reflect real e-commerce payment events.

Dataset Structure

Feature Groups

Temporal

• transaction_ts — parsed UTC timestamp
• hour_of_day — 0–23
• day_of_week — 0–6 (Monday = 0)

Transaction

• transaction_amt — transaction amount in USD
• log_amt — log-transformed amount (ln(amt + 1))
• product_cd — product category (W / H / C / S / R)

Card

• card1 – card6 — card metadata (issuer, type, network)

Identity

• has_identity — 1 if device fingerprint record exists
• id_01 – id_20 — device and browser identity features
• device_type — mobile or desktop

Behavioral (C-features)

• C1 – C14 — counting features (e.g. addresses, payment methods linked to card)

Time-delta (D-features)

• D1 – D15 — days since various reference events

Match flags (M-features)

• M1_enc – M9_enc — encoded boolean match indicators (0/1)

Engineered — Card Velocity

• card1_txn_count — total transactions on this card number
• card1_avg_amt — historical average transaction amount for this card
• card1_historical_fraud_rate — prior fraud rate for this card number

Engineered — Email Risk

• email_txn_count — total transactions from this email domain
• email_historical_fraud_rate — prior fraud rate for this email domain

Engineered — Anomaly Signals

• amt_vs_card_avg_ratio — current amount ÷ card's historical average
• is_high_risk_product — 1 if product_cd = 'W' (highest-fraud product category)

Label

• is_fraud — 0 = legitimate, 1 = fraudulent

Data Splits

The dataset does not include a predefined train/test split. For time-series-safe evaluation, sort by transaction_ts and use a temporal holdout (e.g. last 20%) rather than random split.

Pipeline

Built with:

• DuckDB — columnar in-process ingestion and storage
• dbt — SQL-based transformation layer (staging → marts)
• Python — feature engineering, null handling, timestamp parsing

Full pipeline code: github.com/Kshitijbhatt1998/fintech-fraud-pipeline

Benchmark

A baseline XGBoost model trained on this feature set achieves:

Metric	Score
CV AUC (5-fold time-series)	0.9791
Holdout AUC (last 20% of data)	0.9791
CV Average Precision	see training logs

Training code: src/train.py in the linked repository.

Licensing and Usage

The underlying data is from the Kaggle IEEE-CIS competition. The pipeline code and engineered features are released under Apache 2.0.

This dataset card and the pipeline are provided as a public proof-of-work case study demonstrating custom data pipeline development for fintech AI teams.

Citation

@misc{bhatt2024-fraud-pipeline,
  author = {Bhatt, Kshitij},
  title = {IEEE-CIS Fraud Detection: Model-Ready Feature Dataset},
  year = {2024},
  url = {https://github.com/Kshitijbhatt1998/fintech-fraud-pipeline}
}

Dataset Card Author

Kshitij Bhatt — Data Engineer specializing in fintech AI infrastructure and custom data pipelines.

GitHub | LinkedIn