1

I'm using AWS Personalize to recommend some products. There are 2 test users with totally different impressions data. Yet they are getting the same Recommendations (duplicate). Anyone come across this issue? The 2 Users in this case have very different product impressions, so there should not be any overlap.

Thanks!

kierandes
  • 161
  • 3
  • 16
  • I believe this may be that its reverting to a cold-start recommendation set for some users. Its very strange as some of these users have over 45-50 interactions. Still looking for help here. – kierandes Aug 04 '22 at 15:27
  • 1
    Please clarify what you mean by "impressions". Are these really interactions or truly impressions as Personalize defines them? Note that impressions are used for cold item exploration and indicate items the user has NOT interacted with so they do not drive relevance. – James J Aug 05 '22 at 16:03
  • True, to clarify they are interactions (e.g. users liking an item). – kierandes Aug 08 '22 at 13:42
  • More info is needed. Please add to your question: What recipe are you using? What is your solution configuration (when you created the solution)? What are the offline metrics from the solution version? Composition of your datasets (how many interactions, how many users, how many items). – James J Aug 09 '22 at 18:56

1 Answers1

0

I had the same issue in the past. When recommendations are being the same for different users, then the issue is that your interactions events are being ignored/not recognised correctly.

AWS Personalize has a fixed recommendations set for users, that haven't interacted with the items dataset yet. You can test it yourself, by trying to get recommendations for non-existing user (generate uuid for example). Recommended items will be the same as for users, that you spotted the issue for.

In our case, what we find out, was that all of the interactions were not correctly recognised by Personalize. If you are using PutEvents API to save events in realtime, it might fail silently, if you set sessionId to 'null', instead of omitting this parameter. You will get 200 OK response from Personalize API, however event won't be saved.

You can find out more about this in this GitHub issue, which was indeed closed and not resolved: https://github.com/aws/aws-sdk-js/issues/3371

If that's not the case for you, then I would strongly advise to take a closer look into other PutEvents params, since they could be also causing similar behaviour. Try to experiment with minimal PutEvents API call and be sure that it works (recommendations should be changing immediately for this user). When you are sure, that your API call succeeds, then try adding more stuff from the call, that you are using normally, until it breaks, so you would know, which exactly parameter is causing problems.

PatrykMilewski
  • 922
  • 8
  • 17