Question about User_Personalization_v2 Campaign Behavior with Fake User IDs

Question

Hi Everyone,

I created a campaign using the `user_personalization_v2` recipe. Out of curiosity, I tested it by inputting fake user IDs like "Test" or "John Doe", which I confirmed do not exist in our dataset. I noticed: The campaign still returns results, as if there’s a default recommendation. Could someone provide insights into how these results are determined?

**Additional Notes:**
Using the same dataset group, I created a campaign with the `popularity_count` recipe and tested it with the same fake IDs. The top 11 items were identical. From items 12 to 25, 7 out of 14 were the same (in a different order), while the other 7 were different. This suggests that the `user_personalization_v2` campaign does not derive its results entirely from `popularity_count`.

Thanks!

James · Accepted Answer

Recommendations are provided by the User-Personalization recipes for unknown/cold users as well as known users. From the [Personalize cheat sheet](https://github.com/aws-samples/amazon-personalize-samples/blob/master/PersonalizeCheatSheet2.0.md):

> New users (aka cold users) can be added to your existing User-Personalization solutions by leveraging the PutEvents feature. Each new user starts with a representation in the service that returns the popular items. This representation is shifted by the user’s behavior. As they interact with content within the application and the events are sent by the application to Personalize, recommendations are updated without having to re-train the model. This delivers up to date personalization without constant retraining.

The User-Personalization (v1 and v2) and Popularity-Count recipes use completely different algorithms under the hood. Therefore, recommendations for cold users from these recipes may be similar but are rarely identical. Popularity-Count is simply a counting recipe based on interactions only.

Question about User_Personalization_v2 Campaign Behavior with Fake User IDs

Add your answer

Relevant content