The client was processing more than a half a million transaction per day
Some of these transactions were fraudulent as reported by the card user
The mandate was to predict such transaction and block them
The prediction had to work in a fraction of a second, so that the user experience would not be compromised.
As the number of fraudulent transactions were very low (0.004%), it was very hard to arrive at a generalized solution
Solution
Our solution was an AI based model that provided a confidence score for each transaction. A low score would mean that the transaction is suspicious. The solution included the following capabilities
Sampling of data to reduce the data skew
Feature engineering to add more decision parameters for the AI model
Synthetic data extrapolation
AI based transaction scoring model
Benefits
The client used this model along with a combination of other models that they had to make decisions on the card transactions