credit card fraud detection

Fraud detection data has one defining property: the positive class is vanishingly small. The ULB benchmark contains 284,807 European cardholder transactions over two days in September 2013 with exactly 492 frauds - 0.172% - expressed as 28 PCA-transformed features plus Time, Amount and a Class label. The IEEE-CIS competition set from Vesta scales up: a 590,540-row transaction table with 394 features joined by TransactionID to a 144,233-row identity/device table, about 1.4 GB of CSVs.

Aggregate reporting supplies the third shape. The Reserve Bank of India publishes Part V domestic payment fraud series monthly (shown from September 2022) inside Payment System Indicators, roughly 100 data rows per edition across Parts I-V. Row-level labels train models; national aggregates calibrate expectations about loss rates.

Fraud labels are expensive to produce and heavily constrained once produced. Buyers who skip the fine print discover the limits after their pipeline depends on the data.

• Class imbalance dominates method choice. At 0.172% positives, accuracy is meaningless; evaluation must be recall-at-precision or PR-AUC from day one.
• Anonymization cuts both ways. PCA-transformed features make the ULB set safe to share but impossible to explain to a regulator or a declined customer.
• Competition licenses are narrow. Vesta data restricts use to the competition/research context and prohibits redistribution outside Kaggle without permission.
• Aggregates cannot train models. RBI's monthly fraud totals describe the market; they contain no transaction-level signal whatsoever.

1. Count the positives yourself. ULB documents 492 frauds among 284,807 rows (0.172%); recompute the ratio on load - a file whose class balance differs from spec has been subsampled.

1. Check which columns survived anonymization. The ULB set offers 28 PCA components plus Time, Amount and Class; if your use case needs merchant category or country, this dataset cannot provide it.

1. Verify join keys before planning enrichment. IEEE-CIS links 590,540 transaction rows to 144,233 identity rows via TransactionID - confirm the join cardinality matches your feature design.

1. Read the license against your deployment. Vesta/IEEE-CIS terms confine use to competition and research contexts; ULB carries commercial delivery terms-style listing text with attribution to ULB MLG expected.

1. Pair benchmarks with official aggregates. Use RBI's Part V series (monthly since September 2022) to sanity-check whether your model's alert rates resemble real-world domestic fraud volumes.

Entries that sit next to credit card fraud detection in this glossary:

• fraud-label