Page tree
Skip to end of metadata
Go to start of metadata

Every time a visitor views a promotion on your website, Interaction Studio captures the visitor's "context," including whether he or she is a returning visitor, the device being used, and other information that will give insight into that unique individual. Einstein Decisions, Interaction Studio’s machine learning approach for next best offer decisioning, uses this context to predict the expected value of showing a specific offer to a particular individual by evaluating both the chance of completion as well as the business value of the offer to the company. Using these predictions, Einstein Decisions then determines what promotion or promotions to show to a visitor in order to achieve the highest expected value. Einstein Decisions is an example of a contextual bandit algorithm.

Einstein Decisions automates the process of deciding who should see what content, rather than relying on the manual creation of complex rules. For example, let’s say your financial services company has defined an area on the homepage for highlighting promotions. But with creative for five different offers – credit cards, mortgages, auto loans, checking accounts, and retirement plans – you're not sure which one to display to each site visitor. Einstein Decisions takes away the guesswork by continuously learning from the data points collected about visitors and presents the offer with the highest likelihood of generating the most lift.


IMPORTANT NOTE

Einstein Decisions is available to Interaction Studio Premium Edition customers.

This Article Explains

This article provides an introduction to Einstein Decisions and outlines some of the platform limits that are currently in place.

Sections in this Article

How Does Einstein Decisions Work?

While Einstein Decisions can be combined with prioritization and eligibility rules, it allows for a reduction in the number and complexity of rules that need to be set by leveraging machine-learning to find the maximally relevant promotion for an individual. Einstein Decisions uses its machine-learning capabilities to accomplish this by: 

  • Factoring in extensive data - Einstein Decisions factors in an expansive set of data when making decisions. While it’s always helpful to have as much information as possible about a particular visitor, Einstein Decisions functions effectively even when very little customer data is available. In addition to individual affinities and intent, which may not be known for a first-time visitor, the algorithm also considers information like time-of-day and day-of-the-week along with visitor-specific data such as referring source, device type, browser, geolocation, and time since last visit.
  • Having a simple workflow - within Interaction Studio, you simply add your promotions and their associated assets, assign content zones and tags to the assets, and add any eligibility rules needed. Once deployed, Einstein Decisions uses continuous learning to calculate and present the best experience to each visitor.
  • Being complementary to recommendations - This solution is considered complementary to, rather than a replacement for, Einstein-powered recommendations. While recommendations focus on driving engagement and discovery by presenting products, content, creative, or other catalog aspects like brands, categories, and styles based on an individual’s affinities, Einstein Decisions determines the optimal experience to show to someone based on individual data and the business value to your company.


How Does Einstein Decisions Learn?

Einstein Decisions learns through experience. It is a constant cycle where it will show a promotion, observe any result (configured as a ‘training target’), and update its models to reflect that new observation. Unlike recommendations, which are trying to find the most relevant product or item for a user, Einstein Decisions is trying to identify the promotion which is most likely to obtain the desired result. This means you are directly optimizing for business value as you define it, but there are several caveats you should consider, especially if you are used to working with recommender systems.

  1. A brand new Einstein Decisions implementation starts with a blank slate, and will initially return random results until its first training has completed, which will happen after 100 decisions have been made. Note that adding new promotions to an existing implementation will be used right away, especially if metadata has been used (see the section on adding promotions to your dataset in Configuring Einstein Decisions).
  2. In order to ensure robust learning, Einstein Decisions will sometimes return a random result. This is essential to its design, and is a defining characteristic of Contextual Bandits. It is known as explore vs exploit tradeoff.
  3. If there is no feedback on the training target, learning will never occur and results will continue to be random. For example, if you set the training target to a goal but no user ever completes that goal, Einstein Decisions will not be able to learn.

When Einstein Decisions goes to return a promotion, it does so by making a prediction. The internal machine learning asks “what is the likelihood for each of these promotions that showing them to the user will cause the desired outcome?” and then returns the promotion or promotions with the highest likelihood.

Einstein Decisions Limits

  • Feature Engineering: Up to 200 Features are allowed to be configured on the feature engineering screen
  • Real-Time Promotion Analysis and Ranking Limited to 50 Offers: In order to evaluate and return a next best offer decisions in less than 30ms, Einstein Decisions is limited to evaluating 50 offers at run time. If a user is eligible for more than 50 offers, a randomized set of the eligible offers will be selected for analysis.
  • Promotions in a Campaign Response: A campaign leveraging Einstein Decisions can return up to 12 promotions in a response