Posted On: Jan 27, 2022

We are excited to announce the launch of prediction explanations for Amazon Fraud Detector machine learning (ML) models, available via both the AWS Console and SDK. Prediction explanations report the impact of the predictors (or input variables) on a fraud score, which helps customers achieve greater visibility into how an ML model arrived at a particular fraud score. Amazon Fraud Detector (AFD) 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 ML under the hood and based on over 20 years of fraud detection expertise from AFD automatically identifies potentially fraudulent activity in milliseconds—with no ML expertise required.

Previously, customers received risk scores as part of fraud predictions, but did not get any details describing which of the input variables contributed to a specific ML risk score. This made it difficult to determine how risk scores were calculated, and explain a risk score’s significant contributors for manual investigations, compliance or other purposes. While AFD provides model-level explanations so customers can gain insights into what inputs drive overall model performance, customers were still unable to obtain individual prediction-level explanations.

With prediction explanations, each fraud prediction now comes with information on the impact that each input variable had on a fraud prediction score. These details will help investigators more easily and accurately determine which inputs drove the fraud prediction score up or down. Prediction explanations are included with every prediction at no additional cost.

Customers can view prediction explanations in the AWS Console by navigating to the Fraud Detector console, and clicking on a prediction in the Search Past Predictions tab. Along with each ML-based fraud prediction risk score, a list of the prediction’s event input variables ranked by their impact on the risk score is provided. Customers also get visual indicators of the variable’s significance in terms of magnitude (on a scale of 0 to 5, with 5 being the highest impact on the overall score) and direction (drove the score higher or lower). For example, if the IP address for a given event was the variable that most increased the risk score predicted by the model, it will be listed under 'variables that increased fraud risk' and have a high impact value. Prediction explanations are also available via the AWS SDK and CLI using AFD’s GetEventPredictionMetadata API, making it easy for customers to surface these details to fraud analysts in their preferred investigation workbench.

Prediction explanations are automatically generated and available only for models trained on or after June 30, 2021 in all AWS regions where Fraud Detector is available: US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), Asia Pacific (Singapore) and Asia Pacific (Sydney). For additional details, see our documentation page.