Page tree
Skip to end of metadata
Go to start of metadata
  • Help Content has moved. Help documentation for Marketing Cloud Personalization is now published in Salesforce Help. Starting February 1, 2023, the site will no longer be available.
  • New Name: Interaction Studio (formerly Evergage) is now known as Marketing Cloud Personalization. The new name reflects our mission and vision for innovation in Salesforce Marketing Cloud. We wish we could snap our fingers to update the name everywhere, but you can expect to see the previous name in various places until we replace it.  

Einstein Decisions is easily configured directly within the platform. This article outlines the various components necessary to set-up Einstein Decisions to leverage in cross-channel campaigns.

This Article Explains

In this article, we will outline the various components to setting up and utilizing Interaction Sudio’s next best offer system powered by Einstein Decisions.

Sections in this Article

Einstein Decisions analyzes many unique data points when determining the right experience to present to a particular visitor, whether they are known or anonymous. These are based on visitor behaviors as well as contextual data points like, device type, time of day, and demographic data. In addition to the data points listed in the feature engineering section below, Einstein Decisions can reference additional catalog objects you create in your catalog along with segment membership and user attribute data. To determine what data Einstein Decisions references, you simply define the machine learning features on the Feature Engineering Screen.

Feature Engineering

This screen is where you determine which data features will be captured and trained against by the Einstein Decisions algorithm. A ‘feature’ in machine learning is a data type that we can observe and include in our training. ‘Feature Engineering’ refers to the process of selecting features to be used in a machine learning model. 

Selecting the features for Einstein Decisions to leverage is typically handled by a data science or analytics resource. To configure the Feature Engineering screen, you must have the appropriate user access level and will need to follow these steps:

  1. Login to Interaction Studio
  2. In Machine Learning, select Einstein Decisions > Feature Engineering
  3. Select/Define the data points for the model to consider while training
  4. Select your model target (Options include Conversion, Click, or Goal Completion). This is the action that the model will then try to optimize towards
  5. Click SAVE to save changes

Most options on this screen include a checkbox to determine if they should be considered by Einstein Decisions. Some data points offer additional configuration options. A quick overview of these additional data points are as follows:

  • Segment Membership - select the user segments that you have configured in your dataset that may influence machine learning models for next best offer decisioning 
  • Geography - select the types of geography data you want to capture (e.g. US ZIP code, city, metro area, state/region, country, latitude-longitude, ISP, company)
  • Custom Attributes - what custom user attributes you want to count as features for this machine learning model (e.g. number of support tickets sent in, has enterprise edition, your model scores, loyalty program member, etc)
  • Catalog Object Interaction History - what items in the catalog a user has interacted with, summarized by related catalog objects

Feature Engineering Pro Tips

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 can be used by the algorithm to make better decisions.

What makes a good feature?

Generally, a good feature is any piece of information you think will be relevant to the training target or the promotions themselves. For example, if you have promotions available for winter apparel, including a feature about a shopper's location is a good way to make sure you aren't promoting parkas to shoppers in Florida.

Sometimes features may be included even if they are not useful for personalization because they help identify customers who are likely to reach the training target regardless of what the algorithm does. An example of this is lifetime value. Generally, a shopper who has purchased before is more likely to purchase again, and this helps the machine learning model understand when its decisions made an impact.

What if I include a feature that isn't relevant to my use case?

Einstein Decisions uses algorithms that are robust to irrelevant data. This means that it will learn to ignore features that are not useful. While we don't recommend adding features that are known to be completely irrelevant, it will not significantly hurt the performance of the machine learning model.

Make it simple for me, what features should I include?

Generally speaking, we recommend including all of the default features, as well as adding any custom attributes, segment memberships, and catalog objctcatalog objects that you feel might be relevant. You are limited to 200 features.

Training Target

On the same screen that you select the features for the model to reference, you will also be able to select a training target. The training target is the value that Einstein Decisions is trying to optimize. This is the response that the machine learning training process will observe, and then create a model to maximize. Currently there are three available training targets to choose from:

  1. Click - this action captures when a customer clicks the promotion shown
  2. Conversion - A conversion event is represented by the purchase action in Interaction Studio. Depending on how the client has implemented Interaction Studio, a purchase event could represent a wide variety of actions. For example, a conversion could be a classic ecommerce transaction, or in financial services, it might represent an application submit or resource download. For B2B, a conversion event might be mapped to a webinar signup or account creation. For most retail use cases, conversions will be the best training target to use. The Conversion target uses a 24 hour attribution window from the time a promotion is shown. 
  3. Goal segment - goal segments can be selected when the behavior you want to influence is not well captured by conversions. For example, you may want to optimize the completion of an insurance quote, or requests to contact a salesperson. More information on how to configure a goal segment is available here.

Training targets can be changed at any time, and data that is already collected will be updated to use the new training target, so nothing is lost. It may take up to a day before a change in training target takes effect (the current training target will be leveraged until the new training target is ready). Keep in mind that if you change the training target, it will impact ALL campaigns that are leveraging Einstein Decisions for next best offer decisioning.

Training Target Pro Tips

  • Select the training target that is closest to the business value you are trying to capture. 
  • Avoid using the click training target if the conversion or goal segment target is applicable for your use case. Driving more clicks can be satisfying, but they generally are not inherently valuable by themselves. In some cases, optimizing for click can actually reduce business value, as in the case of click-bait promotions which don't offer a good path to conversion and can end up being over-selected.
  • Even if you do not select clicks as a training target, Einstein Decisions will still record them and learn from them for other training targets. This is useful because even though clicks do not directly represent true business value, they are still a powerful form of direct feedback, and are often more plentiful than conversions or most goal completions.

Adding Promotions to Your Dataset

Once you have configured the feature engineering screen and selected your training target, you need to load in the promotions and their associated assets for Einstein Decisions to choose from. Please see the Promotions article for more information and specific instructions on how to load promotions either directly via the UI or via the Promotion ETL.

Promotion Pro Tips For Effective ML Driven Decisioning

How many promotions should I include?

The more promotions you include, the more likely there will exist a perfect option for each user will exist; however the fewer promotions you include, the quicker learning will occur. Generally speaking, you should expect 10,000 impressions per promotion before you start to see lift from Einstein Decisions, although this could be more or less depending on your use case.

Einstein Decisions uses powerful machine learning to make its assessments. In order to maintain a low latency, it is limited to analyzing and choosing between 50 promotions at a time. You can include more promotions than that in the system, but should use eligibility rules to keep the total number of promotions a single user is eligible for below 50. If more than 50 promotions are eligible for a single user, a random selection of 50 will be chosen for Einstein Decisions to evaluate.

Do I need to include metadata on my promotions?

Metadata is used to give the machine learning model hints about similarities between promotions. For use cases with a small number of promotions, the value will be minimal as the algorithm will learn everything it needs to from user behavior. In use cases where there are a large number of promotions (reduced down to a smaller number of candidates for each user, see above) or when promotions are frequently cycled in and out, metadata can help speed up learning significantly.

Attribute metadata lets you specify properties of your promotion. This might be something like “discount level”, which would give a hint to the machine learning model when comparing promotions that some are offering discounts. The model training will learn faster because it will be able to use what it learned about previous promotions offering a discount when trying to figure out what to do with new promotions which also offer a discount.

In addition to promotion attributes, you can also add related catalog objects much like how you would to any other item in the catalog. Having shared related catalog objects that stretch across multiple catalog objects can help inform Interaction Studio about a customers affinity across items even without direct interaction with each individual object. When the related catalog object on promotions line up with those recorded through Catalog Object Interaction History, as configured on the Feature Engineering screen, learning is greatly improved. When this happens, the machine learning model is able to pay special attention to the user's interaction history for the related catalog object specified on the promotion.

You can safely mix promotions with and without metadata. In this case the model will learn the promotions with metadata quicker and be able to serve them with greater accuracy, but the promotions without metadata will not be penalized.

Can I change promotions after Einstein Decisions has started using them?

It is not recommended to make changes to promotion content after Einstein Decisions has started running with that promotion. The reason is that if Einstein Decisions has already started learning about how users react to a promotion image, for example, and it is replaced with another image, Einstein will take a long time to unlearn what it has already figured out before it can learn how visitors react to the new image.

Promotion metadata, such as attributes and related catalog objects, on the other hand, can be updated if you believe new metadata better describes the promotion. This will be applied retroactively at the next regular training.

What is the minimum number of eligible promotions required for Bandit to function properly?

In order for the Einstein Machine learning training to work properly, there must be more eligible promotions for a decision than the number of decisions returned in a response. That is to say, if you want an Einstein Decisions campaign to present 3 promotions on a home page, you must have at least 4 eligible promotions to choose from in order for learning to occur. If this is not the case, the available promotions will still be shown but there will be no learning since all promotion options were ultimately returned.

Does Einstein Decisions Support Offer/Asset Sorting?

No, sorting is not supported. Einstein Decisions can be leveraged to recommend offers/assets, but does not provide the ability to sort from a set of n items. For example, If a customer has 10 offers that they want to sort and display all 10 on a page, Einstein Decisions should not be used. If a customer wants to determine the best 5 offers to show from a set of 10 offers, this is a supported use case. Einstein Decisions would determine and display the best 5 offers from that set of 10 for each individual customer. There must be at least one more eligible offer than there are slots for display.

Use Einstein Decisions with Promotions

Once you have added promotions to your dataset, you are ready to leverage Einstein Decisions in a campaign. Instructions on how to build a campaign using Einstein Decisions are available at the following links:

Information on how to best interpret Einstein Decisions training results via Einstein Reports is also available here.