How quickly do the events from putEvents influence recommendations ?

Hi-

If using the User-Personalization or Personalized-Ranking recipe, new events will typically impact recommendations within 1-2 seconds for each user. That is, after calling PutEvents, subsequent calls to GetRecommendations/GetPersonalizationRanking (after a couple seconds) for the same user will potentially result in updated recommendations. This also applies to the legacy HRNN-* recipes.

For the User-Personalization recipe, the auto-update process that runs every two hours will bring in new items and impression data added since the last update (for cold item exploration). The auto-update process does not retrain the model.

You still want to periodically retrain your models to incorporate new events/interactions in your models (i.e., across all users). If you're using batch inference jobs, you must retrain (for all recipes) to incorporate new events in the model. Batch jobs currently do not consider new events since the last full retraining.

The SIMS and Similar-Items recipes require retraining to incorporate new events since the last full training.

Hope this helps.

Hi hsakadude,

I think in short, the answer is: it depends on the recipe you use.

I believe putEvents operations update your model when it retrains again, so for the User-Personalization recipe for example, it automatically retrains itself around every two hours as documented here: https://docs.aws.amazon.com/personalize/latest/dg/native-recipe-new-item-USER_PERSONALIZATION.html

There may be some slight updates immediately depending on your filter expressions, etc, but from what I understand the incremental data is not relevant in recommendations until the model is trained to a new solution version.

Hope this helps

Edited by: law-dwg on Nov 15, 2021 12:20 AM
Hi hsakadude,

Please disregard my message and use the one below. It is better informed