We are using Personalized-Ranking-v2 recipe for recommendations. We see a problem where the results seem stale. Is there any other way to improve this?

Question

We are using Personalized-Ranking-v2 recipe for recommendations. We see a problem where the results seem stale. When you request the api for a userId, majority of the recommendations are repeated. This is quite obvious when the user makes no interactions but simply scrolls through the feed. Is this expected? 
we considered sending impressions data to v2. But v2 doesn't support impressions directly or there is a reliable way to send negative signals (ie tell personalize that a recommendation was viewed but not engaged). Is there any other way to improve this?

re:Post Agent · Answer

The Personalized-Ranking-v2 recipe supports real-time personalization, which means it can update recommendations based on user interactions. However, if users are simply scrolling without interacting, the recommendations may appear repetitive because there's no new interaction data to influence the results.

To improve this situation, you should record real-time events using an event tracker and the PutEvents operation. When your recipe supports real-time personalization, Amazon Personalize uses new recorded event data for existing items within seconds of import. This means as users interact with items, the recommendations will adapt to their evolving interests.

Additionally, you can leverage contextual metadata to increase recommendation relevance. By including contextual information such as device type or time of day when getting recommendations, you can provide more varied and relevant results even when interaction patterns are similar.

To maintain recommendation relevance over time, ensure your Item interactions dataset stays current by continuously recording user behavior. Amazon Personalize will update recommendations based on users' most recent activity as they interact with your catalog.

For the Personalized-Ranking-v2 recipe specifically, the system will automatically update to consider new items for recommendations every two hours without requiring a full retraining. However, the model learns most effectively from actual user interactions, so the key to reducing staleness is ensuring you're capturing and recording user engagement events in real-time.

If you're not already doing so, implement comprehensive event tracking to capture user interactions as they happen. This will provide the fresh data needed to keep recommendations dynamic and personalized, even for users who are actively browsing your content.  
**Sources**  
[Increasing recommendation relevance with contextual metadata - Amazon Personalize](https://docs.aws.amazon.com/personalize/latest/dg/contextual-metadata.html)  
[Use case and recipe features - Amazon Personalize](https://docs.aws.amazon.com/personalize/latest/dg/use-case-recipe-features.html)  
[Recording real-time events to influence recommendations - Amazon Personalize](https://docs.aws.amazon.com/personalize/latest/dg/recording-events.html)  
[Maintaining recommendation relevance - Amazon Personalize](https://docs.aws.amazon.com/personalize/latest/dg/maintaining-relevance.html)

We are using Personalized-Ranking-v2 recipe for recommendations. We see a problem where the results seem stale. Is there any other way to improve this?

Add your answer

Relevant content