Fraud Detection
Fraud is rare and costly, so a model has to find a few bad events in a sea of normal ones. The page starts with a handbook and the objectives of fraud detection, moves to papers and methods such as autoencoders and graphs, then streaming tools, and ends with the data sets and how useful they look.
The same notes are in Anomaly Detection.
The best entry point is a full book. Machine Learning for Credit Card Fraud Detection is the Reproducible Machine Learning for Credit Card Fraud detection practical handbook, starting from its foreword, and its code — Practical Handbook Git is the matching repository. The figure below is from the handbook.
Before any model, the goal has to be stated. Fraud detection objectives. is from Nethone, a vendor whose AI-powered user and risk analysis engine is built to detect fraud; the figure below shows those objectives.
With the objective clear, the literature is the next stop. Awesome fraud papers is a curated list of data mining papers about fraud detection. Because fraud looks like an anomaly, one standard method is reconstruction error: Credit card fraud using an autoencoder in Keras is a notebook and pre-trained model that builds a deep autoencoder in Keras for anomaly detection in credit card transactions data.
The same notes are in AUTOENCODERS.
Fraud also hides in relationships between accounts, which is where graphs come in. Graph fraud papers is a curated list of Graph/Transformer-based fraud, anomaly, and outlier detection papers and resources.
The same notes are in Graph Theory.
A method still has to run on live transactions. Fraud using Flink is the repository for the Advanced Flink Application Patterns series, and its docs are the first case study, a fraud detection system that teaches three streaming patterns: dynamic updates of application logic, dynamic data partitioning controlled at runtime, and low latency alerting based on custom windowing logic. For classic models side by side, Credit card fraud on Kaggle does credit card fraud detection through logistic regression, k-means, and deep learning, and Deep graph for fraud is a deep graph-based toolbox for fraud detection. Close to the same money problem, Transforming Financial Forecasting with Data Science and Machine Learning at Uber is Uber's account of financial forecasting.
Data sets
After the methods, the question is which data can actually test them, and the public sets are weaker than they look.
Credit card fraud detection: everyone hit 99%+, seems too easy.
The chargeback set is the opposite problem: questionable fraud data set, it has no usable features and as a time series it doesn't look too informative. https://www.kaggle.com/dmirandaalves/predict-chargeback-frauds-payment
Deprecated links
These links and images no longer work. The original wording is kept here. A same-resource copy, when one was checked, is used above.
Fraud detection on money pools, using social network & pool size, future optimizing using f-beta. This address no longer opens: https://towardsdatascience.com/frauddetection-f801b781410b
Last updated
Was this helpful?