Posted On: Apr 7, 2021
Amazon Fraud Detector is a fully managed service that makes it easy to identify potentially fraudulent online activities, such as the creation of fake accounts or online payment fraud, using customized machine learning (ML) models. To train an ML model, customers provide a dataset that contains examples of legitimate and fraudulent events related to the business activity they want to evaluate for fraud risk. These fraud datasets are often highly imbalanced. For example, a dataset containing one million past transactions may only include 5,000 fraudulent ones, corresponding to a fraud rate of 0.5%. This imbalance in the training data can lead to lower model performance, which results in the customer capturing less fraud. There are a number of common techniques used to treat imbalanced datasets, but applying them requires ML expertise and the best technique often depends on the characteristics of the particular dataset.
Now, Amazon Fraud Detector automatically optimizes your model training dataset if it is imbalanced. Fraud Detector calculates the fraud rate in your dataset and, if it is less than 5%, the service will down-sample the dataset to achieve the optimal distribution. This improves the model’s fraud capture by up to 21% and facilitates faster and more stable model tuning. Customers no longer need to manually adjust their datasets to account for imbalance, and get the benefit from this improved model performance without needing any ML expertise.
Fraud Detector automatically applies these new sampling techniques to any new model versions you train. To train a model version using the console, sign in to the AWS Management Console, open the Amazon Fraud Detector console here, create a model or navigate to an existing model, then select “Actions” followed by “Train new version”.
The new sampling techniques for Amazon Fraud Detector are available today in the US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), Asia Pacific (Singapore) and Asia Pacific (Sydney) regions. For further details, see our documentation.