A financial company would like to detect fraudulent transactions based on customer spending patterns. Detecting such transactions promptly minimizes losses, increases customer satisfaction, and improves security.
The primary source of data used is transaction-specific information such as the amount, the time of purchase, the vendor, and the historical customer behavior. Situational information like federal holidays is added to account for inflated amounts of purchase during sales.
The data is processed, and relevant features are selected. All features are then encoded and scaled appropriately. Since the number of fraudulent transactions is lesser in comparison to that of legitimate transactions, an anomaly detection algorithm is used.
Autoencoders are used for the purpose of this, wherein the model is trained only based only on legitimate transactions. A reconstruction error is then calculated to identify fraudulent transactions.
The model had an accuracy of 75%. This greatly helped in flagging and further investigating suspicious transactions.
** The client has chosen AIDAS Analytics Office (subscription-based pay-as-you-go) services to run business analytics projects.