Variables

Variables represent data elements that you want to use in a fraud prediction. These variables can be taken from the event dataset that you prepared for training your model, from your Amazon Fraud Detector model's risk score outputs, or from Amazon SageMaker AI models. For more information about variables taken from the event dataset, see Get event dataset requirements using the Data models explorer.

The variables you want to use in your fraud prediction must first be created and then added to the event when creating your event type. Each variable you create must be assigned a datatype, a default value, and optionally a variable type. Amazon Fraud Detector enriches some of the variables that you provide such as IP addresses, bank identification numbers (BINs), and phone numbers, to create additional inputs and boost performance for the models that use these variables.

Data types

Variables must have a data type for the data element that the variable represents and can optionally be assigned one of the predefined Variable types. For variables that are assigned to a variable type, the data type is pre-selected. Possible data types include the following types :

Default value

Variables must have a default value. When Amazon Fraud Detector generates fraud predictions, this default value is used to run a rule or model if Amazon Fraud Detector doesn't receive a value for a variable. Default values you provide must match the selected data type. In the AWS Console, Amazon Fraud Detector assigns the default value of 0 for integers, false for Booleans, 0.0 for floats, and (empty) for strings. You can set a custom default value for any of these data types.

Variable types

When you create a variable, you can optionally assign the variable to a variable type. Variable type represents the common data elements that are used to train models and to generate fraud predictions. Only variables with an associated variable type can be used for model training. As part of the model training process, Amazon Fraud Detector uses the variable type associated with the variable to perform variable enrichments, feature engineering, and risk scoring.

Amazon Fraud Detector has pre-defined the following variable types that can be used to assign to your variables.

Assigning variable to a variable type

If you are planning to use a variable for training your model, it is important that you choose a right variable type to assign to the variable. Incorrect variable type assignment can negatively impact your model performance. It can also become very difficult for you change the assignment later, especially if multiple models and events have used the variable.

You can assign your variable any one of the pre-defined variable types or one of the custom variable types – FREE_FORM_TEXT, CATEGORICAL, or NUMERIC.

Important notes for assigning variables to the right variable types

Variable enrichments

Amazon Fraud Detector enriches some of the raw data elements that you provide such as IP addresses, bank identification numbers (BINs), and phone numbers, to create additional inputs and boost performance for the models that use these data elements. The enrichment helps identify potentially suspicious situations and help the models to capture more fraud.

Phone number enrichment

Amazon Fraud Detector enriches phone number data with additional information that relates to geolocation, the original carrier, and the validity of the phone number. Phone number enrichment is automatically enabled for all the models that are trained on or after December 13, 2021 and have a phone number that includes a country code (+xxx). If you have included phone number variable in your model and have trained it before December 13, 2021, retrain your model so it can take advantage of this enrichment.

We highly recommend that you use the following format for phone number variables to ensure that your data is enriched successfully.

Geolocation enrichment

Starting on February 8, 2022 Amazon Fraud Detector calculates the physical distance between the IP_ADDRESS, BILLING_ZIP, and SHIPPING_ZIP values that you provide for an event. The calculated distances are used as inputs to your fraud detection model.

To enable geolocation enrichment, your event data must include at least two of the three variables: IP_ADDRESS, BILLING_ZIP, or SHIPPING_ZIP. In addition, each BILLING_ZIP and SHIPPING_ZIP value must have a valid BILLING_COUNTRY code and SHIPPING_COUNTRY code respectively. If you have a model that was trained before February 8, 2022 and it includes these variables, you must retrain the model to enable the geolocation enrichment.

If Amazon Fraud Detector can't determine the location that's associated with the IP_ADDRESS, BILLING_ZIP ,or SHIPPING_ZIP values for an event due to the data being not valid, a special placeholder value is used instead. For example, suppose that an event has valid IP_ADDRESS and BILLING_ZIP values, but SHIPPING_ZIP value isn't valid. In this case, enrichment is done only for IP_ADDRESS–> BILLING_ZIP. The enrichment isn't done for IP_ADDRESS–>SHIPPING_ZIP and BILLING_ZIP–>SHIPPING_ZIP . Instead, the placeholder values are used in their place. No matter if geolocation enrichment is enabled for your model or not, the performance of your model doesn't change.

You can opt out of geolocation enrichment by mapping your BILLING_ZIP and SHIPPING_ZIP variables to the CUSTOM_CATEGORICAL variable type. Changing the variable type doesn't affect your model's performance.

Geolocation variable format

We highly recommend that you use the following format for geolocation variables to ensure that your location data is enriched successfully.

Useragent enrichment

If you create the Account Takeover Insights (ATI) model, you must provide a variable of the useragent variable type in your dataset. This variable contains the browser, device, and OS data of a login event. Amazon Fraud Detector enriches the useragent data with additional information such as user_agent_family OS_family, and device_family.