You are here:
Feature Engineering in Einstein Decisions
Feature engineering refers to the process of selecting features to use in a machine learning model. A feature is a data type that the model can observe and include in its training. To determine which data Einstein Decisions references, you define the machine learning features on the Feature Engineering page.
Required Editions
| Available in: Premium edition |
Training Target
A training target is the response that the machine learning training process observes, and then creates a model to maximize. Training targets include the following:
| Target | Description |
|---|---|
| Click | This event captures when a user clicks the displayed promotion. |
| Conversion | A conversion event is represented by the purchase action in Marketing Cloud Personalization. Depending upon your implementation, a purchase event can represent a wide variety of actions. For example, a conversion can be a classic ecommerce transaction, or an application submission or resource download. For B2B, a conversion event can map to a webinar signup or account creation. For most retail use cases, conversions are the best training target to use. The conversion target uses a 24-hour attribution window from the time a promotion displays. |
| Goal completion | Using goal completion as your training target can be useful when the behavior you want to influence isn’t captured well by conversions. For example, your goal can be to optimize the completion of an insurance quote, or the fulfillment of requests to contact a salesperson. For more information, see Filters and Global Goals. |
Select the training target that is closest to the business value you’re trying to capture.
Avoid using the click training target if the conversion or goal completion target is applicable for your use case. Driving more clicks can be satisfying, but they generally aren’t inherently valuable by themselves. In some cases, optimizing for click can reduce business value, as with click-bait promotions that don't offer a good path to conversion and can end up being overselected.
Even if you don’t select clicks as a training target, Einstein Decisions still records them and learns from them for other training targets. Even though clicks don’t directly represent true business value, they’re still a valuable form of direct feedback, and are often more plentiful than conversions or most goal completions.
You can change your training target anytime, and data that’s already collected updates to use the new training target, so nothing is lost. However, if you change the training target, it impacts all campaigns that are using Einstein Decisions for next-best-offer decisioning. It can take up to a day before a change in the training target takes effect.
Features
Einstein Decisions creates a machine learning model from rows of training data, where each row contains several columns. Each of these columns is called a “feature” and is a piece of information that the algorithms can use to make better decisions.
Generally, a good feature is any piece of information you think is relevant to the training target or the promotions. For example, if you have promotions for winter apparel, including a feature about a user's location helps ensure you don’t promote parkas to users in Florida.
You can include features even if they aren't useful for personalization because they help identify users that are likely to reach the training target regardless of what the algorithms do. An example is lifetime value. Generally, a user who has purchased before is more likely to purchase again, and this feature helps the machine learning model understand when its decisions make an impact.
Einstein Decisions includes algorithms that account for irrelevant data so that it learns to ignore features that aren’t useful. Although it doesn’t add value to include features that you know are irrelevant, doing so doesn’t impact the performance of the machine learning model.
Salesforce recommends including all of the default features, as well as any custom attributes, segment memberships, and catalog objects that you think are relevant. For the maximum number of features you can specify, see Einstein Decisions Limits.
