Einstein Recipes: How to Build, Test, and Optimize
(To learn more about what Einstein Recipes are, and when, where, and how to use them, check out Einstein Recipes: Interaction Studio’s Secret Ingredient.)
One of Interaction Studio’s most powerful features is Einstein Recipes. These machine-learning powered algorithms curate content based on the “ingredients” selected by the business user to be returned in the form of a recommendation on site, in emails, or even in a mobile app. When strategically placed, these suggestions will increase cart value and encourage continued engagement.
Building a Recipe
When creating a recipe, it’s best to start with a general idea of what the desired results should be. You should think about the location of where the recommendations will be served, how long the user has been on site, and what information is being stored in Interaction Studio about the user. Additionally, knowing which ingredients require an anchor item is key (Co-Buy, Co-Browse, SmartBundle, and Similar Items.) If these ingredients are used in a single-ingredient recipe on a page where no anchor is defined, no recommendations will be returned. It’s best to include multiple ingredients and check the “Allow missing anchor item for multi ingredient recipes” checkbox.
To build a recipe, at least one ingredient must be selected. The options include:
- Collaborative Filtering
- Most Recently Published
- Soon to Expire
- Trending Products
- Similar Items
While most of those seem self-explanatory, some require more information. Collaborative Filtering, for example, utilizes the Alternating Least Squares algorithm, or in other words, returns items based on what “people like me” view, purchase, buy, etc. Similarly, Co-Browse and Co-Buy return items based on how other users interacted with the selected “anchor item” (items in cart, article on page, etc.).
To get strong results, it’s suggested to add multiple ingredients at varying weights. For a product recommendation row on a cart checkout page, the recipe could have a main ingredient of Co-Buy with a weight of 4, followed by Collaborative Filtering with a weight of 3, and the trending fallback box checked to continue serving recommendations to users without a significant amount of data. This recipe would serve a mix of products that are frequently purchased with the item currently in cart, are often purchased or viewed by people with the user’s same affinities, and if neither of those return enough products, trending products would fill in the gaps.
To tailor recipes even more, Interaction Studio allows for inclusions/exclusions, boosters, and variations. Inclusions/exclusions are great for excluding products that are already in a user’s cart, including products for a specific brand, or excluding content already viewed by the user, to name a few. Boosters allow you to hone in further on the recommended items and assign weights to other aspects of the items recommended. For example, a recipe that uses Collaborative Filtering could also include a category booster to emphasize the user’s category affinity over other behaviors that feed into the recipe. Variations ensure that user’s are served a larger breadth of items by randomizing results, or limiting how many of the same brands, categories, colors, etc. are returned.
After selecting ingredients, it’s prudent to review the data model setup to get an idea of what inclusions/exclusions would help strengthen the recipe. How are things like categories, brands, colors, sizes, topics, etc. stored? For a recommendation row on a product display page, it may make sense to include products of the same brand, or for a blog page, content in the same category.
Boosters and variations can be added after the first round of testing to diversify and enhance initial results if needed.
Testing + Optimizing
Once you’ve added ingredients, inclusions/exclusions, boosters, and variations, the next step is to save and “train” the recipe. Depending on how much data is in your dataset and how complicated your recipe is, this could take anywhere from 10 seconds to a minute. Once trained, the option to “Simulate” appears next to a “Sample Test Group”. Clicking “Simulate” opens a new window with options to add users to the test group and a simulation of what items would be recommended to each user in the test group.
Depending on the intended purpose for the recipe, there are a few ways to add users to a test group. You can add users directly from the segment that will be used for campaign targeting, random users, or simply named vs. anonymous users. I find it most useful to add a few users who are highly engaged, users who have items in their cart, and then a few random users. Test groups can be saved and used again and again so once you have a satisfactory mix, you can save it for future recipe testing.
For recipes that include ingredients that require an anchor, be sure to select a product or article to use for the simulation.
Look at the initial product list returned for each user, and ask yourself the following questions:
- Is there variation?
- Do the recommendations match the user’s affinities, what’s in their cart, what they’ve last viewed, etc?
- Is every user getting a recommendation?
Interaction Studio is real-time. If an edit needs to be made, simply toggle back to the recipe set-up, make the adjustment, then re-simulate. Once published, your recipe can be used in web campaigns, Open-Time Email, or server-side campaigns.
Einstein recipes make machine learning accessible and user-friendly. If you want to learn more about how to make your Einstein Recipes great, contact us!