Recommendation Strategies

In this topic, you'll learn about:

Strategies

Recommendation strategy comprises:

  • An algorithm that produces recommendations.
  • Filters to curate the item/content recommendations based on the item/content attributes.
  • Fallback items that are always recommended in case there are no recommendations that can be presented.

Recommendation models/algorithms

This section explains how each algorithm produces recommendations.

Popularity-based models

1. Most Viewed

Purpose of the Model

This model returns as recommendations items that were most frequently viewed.

Model Logic

Every item's score is calculated as below:

  • Item Score: (Number of times the item was viewed) * (Time factor to account for recency of view)

  • Example: If an item was viewed 100 times on Monday, 200 times on Tuesday, and 300 times on Wednesday, then on Thursday, the Most Viewed model will give this item the score of 598 = 100 * 0.99 + 200 * 0.995 + 300.

Model Output Items are then ordered in descending order of their score, and the top 12 items that are eligible for recommendations and that satisfy filtering conditions are returned in the content generator response. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy the filtering conditions. In this case, all such items are returned in the content generator response.

2. Best Sellers

Purpose of the Model This model returns, as recommendations items that were most frequently bought.
Model Logic

Every item's score is calculated as below:

  • Item Score: (Number of times the Item was purchased) * (Time factor to account for recency of view)

  • Example: If an item was bought 100 times on Monday, 200 times on Tuesday, and 300 times on Wednesday, then Best Seller model built on Thursday will give this item the score of 598 = 100 * 0.99 + 200 * 0.995 + 300.

Model Output Items are then ordered in descending order of their score, and the top 12 items that are eligible for recommendations and that satisfy filtering conditions are returned in the content generator response. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy filtering conditions. In this case, all such items are returned in the content generator response.

3. Trending - views

Purpose of the Model This model returns as recommendations items that were most frequently viewed in the last 7 days.
Model Logic

Every item's score is calculated as below:

  • Item Score: (Number of times the Item was viewed in last 7 days) * (Time factor to account for recency of view)

  • Example: If an item was viewed 100 times on Monday, 200 times on Tuesday, and 300 times on Wednesday, then on Thursday, the Most Viewed model will give this item the score of 598 = 100 * 0.99 + 200 * 0.995 + 300.

Model Output Items are then ordered in descending order of their score, and the top 12 items that are eligible for recommendations and that satisfy filtering conditions are returned in the content generator response. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy the filtering conditions. In this case, all such items are returned in the content generator response.

4. Trending - purchases

Purposes of the Model This model returns as recommendations items that were most frequently bought in the last 7 days.
Model Logic

Every item's score is calculated as below:

  • Item Score: (Number of times the Item was bought in last 7 days) * (Time factor to account for recency of view)

  • Example: If an item was bought 100 times on Monday, 200 times on Tuesday, and 300 times on Wednesday, then Best Seller model built on Thursday will give this item the score of 598 = 100 * 0.99 + 200 * 0.995 + 300.

Model Output Items are then ordered in descending order of their score, and the top 12 items that are eligible for recommendations and that satisfy filtering conditions are returned in the content generator response. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy filtering conditions. In this case, all such items are returned in the content generator response.

Co-occurrence based models

1. Viewed this, Viewed that

Purpose of the Model This model returns as recommendations items that were most frequently viewed together with the item being currently viewed (the ID of this item should be passed in the recommendations request to the content generator).
Model Logic For every item, its score is calculated depending on the number of users who viewed both this item and the currently viewed item.
Model Output Items are then ordered in descending order of their score, and the top 23 items are selected. 12 recommendations in the content generator response are first filled in with items from those 23 items that satisfy filtering conditions. The remaining empty slots are then filled with recommendations built by the Most Viewed model that satisfy filtering conditions and do not duplicate already selected items. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy filtering conditions. In this case, all such items are returned in the content generator response. The Viewed this, Viewed that model is designed to work on Product Detail Pages (PDPs) where the item being currently viewed is present and its ID can be passed to the model as a parameter. Note that if the model receives an ID that is not present in the inventory file or does not receive the ID of the currently viewed item, then it returns recommendations built with the help of the Most Viewed model.

2. Viewed this, Bought that

Purpose of the Model This model returns as recommendations items that were most frequently bought by users who also viewed the currently viewed item (the ID of this item should be passed in the recommendations request to the content generator).
Model Logic For every item, its score is calculated depending on the number of users who both bought this item and viewed the currently viewed item.
Model Output Items are then ordered in descending order of their score, and the top 23 items are selected. 12 recommendations in the content generator response are first filled with items from those 23 items that satisfy filtering conditions. The remaining empty slots are then filled in with recommendations built by the Most Viewed model that satisfy filtering conditions and do not duplicate already selected items. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy filtering conditions. In this case, all such items are returned in the content generator response.

The Viewed this, Bought that model is designed to work on PDPs where the item being currently viewed is present and its ID can be passed to the model as a parameter. Note that if the model receives an ID that is not present in the inventory file or does not receive the ID of the currently viewed item, then it returns recommendations built with the help of the Most Viewed model.

3. Bought this, Bought that

Purpose of the Model This model returns as recommendations items that were most frequently bought together with the currently viewed item (the ID of this item should be passed in the recommendations request to the content generator).
Model Logic For every item, its score is calculated depending on the number of users who bought both this item and the currently viewed item.
Model Output

Items are then ordered in descending order of their score, and the top 23 items are selected. 12 recommendations in the content generator response are first filled with those 23 items that satisfy filtering conditions. The remaining empty slots are then filled with recommendations built by the Best Sellers model that satisfy filtering conditions and do not duplicate already selected items. The model may produce less than 12 recommendations if there are less than 12 items in the inventory file that are eligible for recommendations and that satisfy filtering conditions. In this case, all such items are returned in the content generator response.

The Bought this, Bought that model is designed to work on PDPs where the currently viewed item is present and its ID can be passed to the model as a parameter. Note that if the model receives the ID that is not present in the inventory file or does not receive the ID of the currently viewed item, then it returns recommendations built with the help of the Best Sellers model.

User-history based models

1. Last Viewed

Note: This model ignores any configured filtering conditions.

Purpose of the Model This model returns as recommendations 9 items that were most recently interacted with, by the user.
Model Logic Interaction is classified as a view or purchase of an item. Items in recommendations are ordered by their interaction recency, with the most recently interacted item holding the first position.
Model Output The model can return up to 9 recommendations in the content generator response, but no recommendations are returned for a user who has not previously interacted with any item.

2. 1-1 Recommendations

Purpose of the Model

For a user with history of interactions with items on the client website, the algorithm produces recommendations based on user history and some similarity between items that is calculated based on past user-item interactions on the website. More recent items in user history have higher importance when generating recommendations.

Model Logic

This algorithm belongs to the family of collaborative filtering approaches. Its hyper-parameters are selected via regularly performed offline validation procedure with the goal to maximize the chances of predicting the next item the user will interact with. This ensures the best algorithm is selected for the client.

To configure 1-1 Recommendations for email for retargeting use cases, you can do either of the following:

External Recommendations

Note: This model ignores any configured filtering conditions and workspace-level fallback models.

Purpose of the Model

This model allows the use of Recommendations REST API, as well as email and web widgets, to display recommendations generated by an external system.

Model Logic

The model does not utilize a machine-learning back-end of Recommendations. Instead, it accepts a list of IDs generated by an external recommendation provider and returns the same items, but with the associated data available for those items in the Inventory database.

Model Output

Unlike other models, this one is designed to maintain a fixed slot for each recommendation, regardless of whether some items are missing from the Inventory. You can set a predefined list of fallback items so that if a recommended product has no associated record in the inventory, it will be replaced by a fallback item in the same slot. For example, if the external system generates items 1, 2, and 3, but there is no inventory data for item 2, a fallback item will appear in slot 2 while item 3 remains in slot 3.

Fallback items

Fallback items are standard or commonly recommended items preselected by the marketers. These items are displayed when recommendation filter criteria does not yield any results. This option is very useful and kicks in when customers don't yet have any interaction (views or purchase) history, or when there are no items that match the marketer's campaign recommendation filter criteria.

This option will ensure some items/content are always displayed in the recommendations widget even when the algorithms along with the filters applied do not yield any results. These can be entered as individual items in the strategy.

For example, let us say we are using a most purchased model, which uses the user’s purchase history to make recommendations of similar items that they may purchase. What happens in the scenario when there is no data available to build the recommendations? This could be the case if it is the user's first visit to the site and has not purchased any items yet.

In such a scenario, it is important that there is a fallback mechanism in place, and we can show the user some item recommendations. The marketer could set these or could use a most popular model for the fallback items, so that the space occupied by the recommendations could result in a customer interaction and not go to waste.

Learn more

Setting Up Recommendations

Adding Email Recommendations

Creating Recommendations Service and Getting Recommendations Using REST API

Adding Recommendations to a Website Using Web Personalization

Recommendations Filtering