Input and output in Customer AI
The following document outlines the different required events, inputs, and outputs utilized in Customer AI.
Getting started getting-started
Here are the steps to build propensity models and identify target audiences for personalized marketing in Customer AI:
-
Outline use cases: How would propensity models help to identify target audiences for personalized marketing? What are my business goals and corresponding tactics to achieve the goal? Where can propensity modeling fit in this process?
-
Prioritize use cases: Which are the highest priorities for the business?
-
Build models in Customer AI: Watch this quick tutorial and refer to our UI guide for a step-by-step process to build a model.
-
Build segments using model results.
-
Take targeted business actions based on these segments. Monitor the results and iterate over the actions to improve.
Here are example configurations for your first model. The example model, built in this document, uses a Customer AI model to predict who is likely to convert for a retail business in the next 30 days. The input dataset is an Adobe Analytics dataset.
Model Type: Conversion
Identity: Ensure the identity column for each dataset is set to be a common identity.
commerce.purchases.value
equals to pencilOutcome window: 30 days.
Enable for profile: This must be enabled for the model output to be used in segmentation.
Data overview data-overview
The following sections outline the different required events, inputs, and outputs utilized in Customer AI.
Customer AI works by analyzing the following datasets to predict churn (when a customer is likely to stop using the product) or conversion (when a customer is likely to make a purchase) propensity scores:
- Adobe Analytics data using the Analytics source connector
- Adobe Audience Manager data using the Audience Manager source connector
- Experience Event dataset
- Consumer Experience Event dataset
You can add multiple datasets from different sources if each of the datasets share the same identity type (namespace) such as an ECID. For more information on adding multiple datasets, visit the Customer AI user guide.
The following table outlines some common terminology used in this document:
Experience Event
. The data behavior of a schema is defined by the schema’s class, which is assigned to a schema when it is first created. XDM classes describe the smallest number of properties a schema must contain in order to represent a particular data behavior.meta:intendedToExtend
attribute.Customer AI input data customer-ai-input-data
For input datasets, like Adobe Analytics and Adobe Audience Manager, the respective source connectors directly map the events in these standard field groups (Commerce, Web, Application, and Search) by default during the connection process. The table below shows the event fields in the default standard field groups for Customer AI.
For more information on mapping Adobe Analytics data or Audience Manager data, visit the Analytics field mappings or Audience Manager field mappings guide.
You can use Experience Event or Consumer Experience Event XDM schemas for input datasets that are not populated via one of the above connectors. Additional XDM field groups can be added during the schema creation process. The field groups can be provided by Adobe like the standard field groups or a custom field group, which matches the data representation in the Platform.
Standard field groups used by Customer AI standard-events
Experience Events are used for determining various customer behaviors. Depending on how your data is structured, the event types listed below may not encompass all of your customer’s behaviors. It is up to you to determine what fields have the necessary data that is needed to identify web or other channel-specific user activity clearly and unambiguously. Depending on your prediction goal, the required fields that are needed can change.
Customer AI uses the events in these four standard field groups by default: Commerce, Web, Application, and Search. It is not necessary to have data for each event in the standard field groups listed below but certain events are required for certain scenarios. If you have any events in the standard field groups available, it is recommended that you include it in your schema. For example, if you wanted to create a Customer AI model for predicting purchase events, it is useful to have data from the Commerce and Web page details field groups.
To view a field group in the Platform UI, select the Schemas tab on the left-rail followed by selecting the Field groups tab.
-
commerce.order.purchaseID
-
productListItems.SKU
-
commerce.productListViews.value
-
productListItems.SKU
-
commerce.checkouts.value
-
productListItems.SKU
-
commerce.purchases.value
-
productListItems.SKU
-
commerce.productListRemovals.value
-
productListItems.SKU
-
commerce.productListOpens.value
-
productListItems.SKU
-
commerce.productViews.value
-
productListItems.SKU
web.webPageDetails.name
web.webInteraction.linkClicks.value
-
application.applicationCloses.value
-
application.name
-
application.crashes.value
-
application.name
-
application.featureUsages.value
-
application.name
-
application.firstLaunches.value
-
application.name
-
application.installs.value
-
application.name
-
application.launches.value
-
application.name
-
application.upgrades.value
-
application.name
search.keywords
Additionally, Customer AI can use subscription data to build better churn models. Subscription data is needed for each profile using the Subscription data type format. Most of the fields are optional, however, for an optimal churn model it is highly recommended that you provide data for as many fields as possible such as, startDate
, endDate
, and any other relevant details. Please reach out to your account team for additional support on this feature.
Adding custom events and profile attributes add-custom-events
If you have information you wish to include in addition to the default standard event fields used by Customer AI, you can use the custom event configuration to augment the data used by the model.
When to use custom events
Custom events are necessary when the datasets chosen in the dataset selection step contain none of the default event fields used by Customer AI. Customer AI needs information about at least one user behavior event other than the outcome.
Custom events are helpful for:
-
Incorporating domain knowledge or prior expertise into the model.
-
Improving the predictive model quality.
-
Gaining additional insights and interpretations.
The best candidates for custom events are data that contain domain knowledge that may be predictive of the outcome. Some general examples of custom events include:
-
Register for account
-
Subscribe to newsletter
-
Make a call to customer service
The following are a selection of industry-specific custom event examples:
Sign up for club card
Clip mobile coupon.
Stream video.
Purchase loyalty points.
Custom events must represent user-initiated actions in order to be selected. For example, “Email Send” is an action initiated by a marketer and not by the user, so it shouldn’t be used as a custom event.
Historical data
Customer AI requires historical data for model training. The required duration for data to exist within the system is determined by two key elements: the outcome window and eligible population.
By default, Customer AI looks for a user to have had activity in the last 45 days if no eligible population definition is provided during the application configuration. Additionally, Customer AI requires a minimum of 500 qualifying and 500 non-qualifying events (1000 total) from historical data based on a predicted goal definition.
The following examples demonstrate the use of a simple formula which helps you determine the minimum amount of data required. If you have more data than the minimum requirement, your model is likely to provide more accurate results. If you have less than the minimum amount required, the model will fail, as there is not enough data for model training.
Formula:
To decide the minimum required duration of data existing within the system:
-
The minimum data required to create features is 30 days. Compare the eligibility lookback window with 30 days:
-
If the eligibility lookback window is greater than 30 days, the data requirement = eligibility lookback window + outcome window.
-
Otherwise, the data requirement = 30 days + outcome window.
-
** If there is more than one condition for defining the eligible population, the eligibility lookback window is the longest one.
Examples:
-
You want to predict whether a customer is likely to purchase a watch in the next 30 days for those who have some web activity in the last 60 days.
-
Eligibility lookback window = 60 days
-
Outcome window = 30 days
-
Data required = 60 days + 30 days = 90 days
-
-
You want to predict whether the user is likely to purchase a watch in the next 7 days without providing an explicit eligible population. In this case, the eligible population defaults to “those who have had activity in the last 45 days” and the outcome window is 7 days.
-
Eligibility lookback window = 45 days
-
Outcome window = 7 days
-
Data required = 45 days + 7 days = 52 days
-
-
You want to predict whether the customer is likely to purchase a watch in the next 7 days for those who have some web activity in the last 7 days.
-
Eligibility lookback window = 7 days
-
Minimum data required to create features = 30 days
-
Outcome window = 7 days
-
Data required = 30 days + 7 days = 37 days
-
Although Customer AI requires a minimum period of time for the data to exist within the system, it also works best with recent data. By using more recent behavioral data, Customer AI is likely to yield a more accurate prediction of a user’s future behavior.
Customer AI output data customer-ai-output-data
Customer AI generates several attributes for individual profiles that are deemed eligible. There are two ways to consume the score (output) based on what you have provisioned. If you have a Real-time Customer Profile-enabled dataset, you can consume insights from Real-time Customer Profile in the Segment Builder. If you don’t have a Profile-enabled dataset, you can download the Customer AI output dataset available on the data lake.
You can find the output dataset in the Platform Datasets workspace. All Customer AI output datasets start with the name Customer AI Scores - NAME_OF_APP. Similarly, all Customer AI output schemas start with the name Customer AI Schema - Name_of_app.
The table below describes the various attributes found in the output of Customer AI:
These are predicted reasons as to why a profile is likely to convert or churn. These factors are comprised of the following attributes:
- Code: The profile or behavioral attribute which positively influences a profile’s predicted score.
- Value: The value of the profile or behavioral attribute.
- Importance: Indicates the weight of the profile or behavioral attribute has on the predicted score (low, medium, high)
Next steps next-steps
Once you prepare your data and ensure that all your credentials and schemas are in place, refer to the Configure a Customer AI Instance guide, which walks you through a step-by-step tutorial to create a Customer AI instance.