Prompt
How do I find a dataset for fraud detection with enough labeled examples?
Latest observation
To find a fraud detection dataset with enough labeled examples, focus on a few practical sources and filter for label quality, class balance, and domain match.
Good places to look
1. Public benchmark repositories
These are the fastest way to get started:
-
Kaggle
Search terms:- “credit card fraud”
- “transaction fraud”
- “insurance fraud”
- “payment fraud”
- “anomaly detection fraud”
Popular datasets often include:
- Credit card transaction fraud
- IEEE-CIS fraud detection
- synthetic fraud datasets
-
UCI Machine Learning Repository Some fraud-related datasets are available, though many are smaller or outdated.
-
OpenML Useful for tabular datasets with metadata and task descriptions.
-
Hugging Face Datasets Less common for fraud, but worth checking for transaction or review fraud tasks.
2. Competition datasets
Older competition datasets can be very useful because they often have:
- labeled target variable
- train/test split
- lots of examples
- feature engineering guidance from discussions
Examples:
- IEEE-CIS Fraud Detection
- Credit Card Fraud Detection
- Porto Seguro’s Safe Driver Prediction isn’t fraud, but is often used as an imbalanced classification benchmark
3. Domain-specific open data
Depending on the fraud type:
- Financial fraud: transaction datasets, AML/anti-money laundering datasets
- Insurance fraud: claims datasets
- E-commerce fraud: payment disputes, chargebacks
- Identity fraud: account takeover or signup abuse datasets
- Review fraud: fake review datasets
4. Synthetic data
If real labeled fraud data is scarce, use:
- synthetic generators
- simulation-based datasets
- fraud injection into normal transaction data
This is common because real fraud labels are often rare and sensitive.
How to judge whether a dataset has “enough” labels
Look for:
-
Number of positive fraud cases
Fraud is usually rare, so total size matters less than the count of fraud examples.
A dataset with 1M rows but only 500 fraud cases may still be hard to learn from. -
Label quality
- Are labels confirmed by investigation or chargeback?
- Are they delayed?
- Are there unlabeled suspicious cases mixed in?
-
Class imbalance Fraud datasets are often highly imbalanced, sometimes 0.1% or less positive.
-
Temporal coverage Fraud behavior changes over time. Prefer datasets with timestamps.
-
Feature realism Check whether features are usable for your problem:
- transaction amount
- device/browser
- merchant/category
- geolocation
- account history
-
Leakage risk Avoid datasets where the label is directly or indirectly encoded in features.
Search strategy that works well
Use combinations like:
site:kaggle.com fraud detection dataset labeledsite:openml.org fraud datasetsite:uci.edu fraud classification datasetcredit card fraud labeled transactionschargeback dataset fraud detectionsynthetic fraud detection dataset
Also search by your exact fraud type:
- card fraud
- insurance fraud
- loan fraud
- account takeover
- AML
- fake reviews
- claims fraud
If you need a lot of labeled examples
If public data is insufficient, consider:
-
Combining multiple datasets
- Use several fraud datasets for pretraining
- Adapt to your domain via fine-tuning
-
Weak supervision
- Use rules, heuristics, and suspiciousness signals to create proxy labels
-
Active learning
- Label only the most informative cases
-
Semi-supervised learning
- Train mostly on normal data, with a small set of fraud labels
-
Vendor or consortium data
- Banks, payment processors, insurers, and fraud platforms often have shared or commercial datasets
Recommended starting datasets
If you want well-known, labeled tabular fraud benchmarks:
- Credit Card Fraud Detection on Kaggle
- IEEE-CIS Fraud Detection on Kaggle
- PaySim synthetic mobile money fraud dataset
- BankSim synthetic banking fraud dataset
Rule of thumb
If your goal is model development, try to get at least:
- hundreds to thousands of fraud cases, not just millions of total rows
- timestamps
- transaction/account-level features
- a clear definition of fraud
If you want, I can also help you:
- find datasets for a specific fraud type, or
- compare a few candidate datasets for label count and suitability.