Maintaining recommendation relevance - Amazon Personalize
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China.

Maintaining recommendation relevance

Maintain the relevance of recommendations to increase user engagement, click-through rate, and conversion rate for your application as your catalogue grows. To maintain and improve the relevance of Amazon Personalize recommendations for your users, keep your data and solution versions up to date. This allows Amazon Personalize to learn from your user’s most recent behavior and include your newest items in recommendations.

You can automate and schedule re-training and data import tasks with Maintaining Personalized Experiences with Machine Learning, an Amazon Solutions Implementation that automates the Amazon Personalize workflow, including data import, solution version training, and batch workflows. For more information see Maintaining Personalized Experiences with Machine Learning.

Keeping datasets current

As your catalog grows, update your historical data with bulk or incremental data import operations. We recommend that you first import your records in bulk, and then incrementally add items and users as your catalog grows. For more information about importing historical data see Preparing and importing data.

For real-time recommendations, keep your Interactions dataset up to date with your users' behavior by recording interaction events with an event tracker and the PutEvents operation. Amazon Personalize updates recommendations based on your user's most recent activity as they interact with your application. For more information on recording real-time events, see Recording events.

If you have already created a solution version (trained a model), new records influence recommendations as follows:

  • For new events, Amazon Personalize immediately uses historical and real-time interaction events between a user and existing items (items you included in the data you used to train the latest model) when generating recommendations for the same user. Historical events that you import using the Amazon Personalize console and events that you record in real-time influence recommendations in the same way. For more information, see How real-time events influence recommendations.

  • For new items, if you trained the solution version with User-Personalization, Amazon Personalize automatically updates the model every two hours. After each update, the new items can be included in recommendations with exploration. For information about exploration see User-Personalization recipe.

    For any other recipe, you must retrain the model for the new items to be included in recommendations.

  • For new users, recommendations will initially be only for popular items. Starting with the first event, user recommendations will be more relevant as you record events. For more information, see Recording events.

Keeping solution versions up to date

Create a new solution version (retrain the model) to include new items in recommendations and update the model with your user’s most recent behavior. Once you retrain the model, you must update the campaign to deploy it. For more information see Updating a campaign.

Retraining frequency depends on your business requirements and the recipe that you use. For most workloads, we recommend training a new model weekly with training mode set to FULL. This creates a completely new solution version based on the entirety of the training data from the datasets in your dataset group. If you add new items frequently and you don't use User-Personalization, you may need to do a full retraining more often to include those new items in recommendations.

With User-Personalization, Amazon Personalize automatically updates your latest fully trained solution version (trained with trainingMode set to FULL) every two hours to include new items in recommendations with exploration. Your campaign automatically uses the updated solution version. This is not a full retraining; you should still train a new solution version weekly with trainingMode set to FULL so the model can learn from your users' behavior.

If every two hours is not frequent enough, you can manually create a solution version with trainingMode set to UPDATE to include those new items in recommendations. Just remember that Amazon Personalize automatically updates only your latest fully trained solution version, so the manually updated solution version won't be automatically updated in the future. For more information, see User-Personalization recipe.

For information on creating a new solution version, see Step 3: Creating a solution version.