How new data influences real-time recommendations
After you create a recommender or custom solution version, how new data influences real-time recommendations depends on its type, the method of import, and the domain use case or custom recipe you use. The following sections explain how new data influences real-time recommendations before the next training.
Training can be
a recommender's weekly automatic training, or automatic or manual solution version creation.
For manual training with User-Personalization, you must set trainingMode to FULL.
For information on how new records influence batch recommendations, see Getting batch recommendations. For information on how new records influence batch segment jobs, see Getting user segments.
New interactions
New interactions are item or action interactions that you import after the latest training.
For both real-time and bulk data, if interactions involve a new item or action, Amazon Personalize might consider it for recommendations without training. For more information, see New items or New actions.
Real-time events
For use cases and recipes that feature real-time personalization, Amazon Personalize immediately uses real-time interactions between a user and existing items or actions (records present at the latest training) when generating recommendations for the same user. For more information about real-time personalization, see Real-time personalization.
For any domain use cases and custom recipes that don't feature real-time personalization, such as recommending similar items, your model learns from real-time interactions data only after training.
Bulk interactions
For bulk interactions, for both incremental and full dataset import jobs, your model learns from bulk item interaction or action interaction data only after the next training. Bulk data isn't used to update recommendations for real-time personalization.
For more information about updating existing bulk data, see Updating existing bulk data.
New items
New items are items that you import after the latest training. They can come from either interactions data or item metadata in an Items dataset.
New items are considered for recommendations as follows:
-
For Top picks for you and Recommended for you domain cases or User-Personalization or Next-Best-Action recipes, Amazon Personalize automatically updates the model every two hours. After each update, Amazon Personalize considers new items for recommendations as part of exploration. When considering the new item, Amazon Personalize considers any metadata for the item. However this data will have a greater effect on recommendations only after you record interactions for the item and train a new model. For information about updates, see Automatic updates.
-
If you use the Trending now use case, Amazon Personalize automatically evaluates your interactions data every two hours and identifies trending items. You don't have to wait for your recommender to train. If you use the Trending-Now recipe, Amazon Personalize automatically considers all new items over configurable intervals without training. For information about configuring intervals, see Trending-Now recipe.
-
If you don't use the Trending-Now recipe or your use case or recipe doesn't support automatic updates, Amazon Personalize will consider new items only after the next training.
New users
New users are users that you import after the latest training.
They can come from either interactions data or user metadata in a Users dataset. For new, anonymous users (users without a userId),
you can record events for the user with a sessionId and Amazon Personalize will
associate events with the user before they log in. For more information see Recording events for
anonymous users.
Amazon Personalize generates recommendations for new users as follows:
-
If you use the Trending now domain use case or Trending-Now custom recipe, new users immediately receive recommendations for the top trending items. If you use the Popularity-Count recipe, new users immediately receive recommendations for items with the most interactions.
-
For recipes or use cases that provide personalized recommendations for users, recommendations for new users are based on the early interaction histories of your existing users. The first items or actions these existing users interacted with are more likely to be recommended to new users. For the User-Personalization or Personalized-Ranking recipes, if you set
recency_masktotrue, recommendations also include items based on the latest popularity trends in your interactions data.
The following can increase recommendation relevance for new users:
-
Interactions data – The primary way to improve recommendation relevance for a new user is to import data from their interactions with your items. For information about how new interactions data influences recommendations, see New interactions.
-
User metadata – Importing user metadata, such as GENDER or MEMBERSHIP_STATUS, can improve recommendations. For metadata to influence recommendations, you must wait for your domain recommender's weekly automatic retraining to complete. Or you must manually create a new solution version.
-
Contextual metadata – If your use case or recipe supports contextual metadata and your Item interactions dataset has metadata fields for contextual data, you can provide the user's context in your request for recommendations. This does not require retraining. For more information, see Increasing recommendation relevance with contextual metadata.
New actions
New actions are actions that you import since the latest training. They can come from either action interaction data or actions in an Actions dataset.
With the Next-Best-Action recipe, Amazon Personalize automatically updates a solution version every two hours. After each update, Amazon Personalize considers new actions for recommendations as part of exploration. When considering the new action, Amazon Personalize considers any metadata for the action. However, this data will have a greater effect on recommendations only after you record action interactions for the action and fully retrain. For information about updates, see Automatic updates