Unit8 Talks #7 - Fraud detection - A guide to building a financial transaction anomaly detector

Unit8 Talks #7 - On technology - Fraud detection - A guide to building a financial transaction anomaly detector
Many companies currently still rely on hard-coded, inflexible ways of detecting anomalies in their financial transactions, hence leading to lots of false positives or risking fraudulent transactions to be executed. How can anomalies be found in financial transactions when we don't know what indicators we are looking for in advance?
One possible solution to this is building a machine learning model that attempts to automate the approach. The model measures how easy individual data points can be separated from the rest of the data (i.e. using an isolation forest).
In this webinar we will explore how to build such a Machine Learning model, how to use its outputs, and how to create a complementary explanation model to interpret and validate these outputs..
Who should attend?
- Technical Executives, Data Scientists, Controlling/Risk teams
- Technical level L200
Why should you attend?
- A guide to machine learning driven fraud detection applicable across different industries - which can be reused to your needs
Get in touch [e-mail]
More about Unit8 - unit8.co/
Unit8 on LinkedIn - / unit8.co
Unit8 on Twitter - / unit8co
Unit8 on Instagram - / unit8.co
Unit8 on Medium - / unit8-machine-learning

[@dhuxdheerdahir2736]

Thank you

[@hakeemojulari7392]

Great talks! I would like to ask if it is possible to use anomaly detection to detect fraud in ATM transactions with the following features: CardNo, branch-code, AtmID, Trans-date, Amount, Trans-type, Trans-status How can the customer's regular transaction patterns be used to detect anomalies (suspicious fraud)?

[@brackly]

Hi @hakeemojulari7392, I dont know if you got an answer to this but its possible. You might want to look into Deep learning architectures called Autoencoders. In theory, if you give an autoencoder the regular customer's data, you can train it to reconstruct every single datapoint. What hapens as a result is that it learns patterns regarding normal transaction behavior, and the reconstruction loss gets smaller and smaller. You can then use the reconstruction loss of the autoencoder to detect suspicious transactions. Intuitively, if a transaction is so new that the autoencoder has never seen before, the reconstruction loss will be higher that a transaction that it has seen before therefore flagging it.

[@mubangalillian2640]

Can it also be said the anomaly detection model can used to label data that can be used in supervised learning model that can be used to for fraud detection?

[@mathematics-in3wi]

What is require feature for anomaly detection?

[@radhika4573]

How to undo onehotencoding and add shap values

[@namanladha9378]

Can you share the code?

[@mayasushma2951]

Please share the code