Good day.

I am using AWS Personalize to generate recommendations and I am trying to upgrade to “aws-user-personalization-v2” from “aws-user-personalization”. The model has been trained with user-item-iteractions which includes a “TIME-OF-DAY” field witch takes values such as: “morning”, “afternoon”, “evening”.

The schema looks like this:

{
    "type": "record",
    "name": "Interactions",
    "namespace": "com.amazonaws.personalize.schema",
    "fields": [
        {
            "name": "USER_ID",
            "type": "string",
            "categorical": false
        },
        {
            "name": "ITEM_ID",
            "type": "string",
            "categorical": false
        },
        {
            "name": "TIMESTAMP",
            "type": "long",
            "categorical": false
        },
        {
            "name": "EVENT_TYPE",
            "type": "string",
            "categorical": true
        },
        {
            "name": "EVENT_VALUE",
            "type": "float",
            "categorical": false
        },
        {
            "name": "TIME_OF_DAY",
            "type": "string",
            "categorical": true
        }
    ],
    "version": "1.0"
}

The Items schema conforms to the default schema as on AWS’s documentation. I query Personalize using boto3:

ACCOUNT = ...
REGION = ...

personalize_client = boto3.client("personalize-runtime")

...

def call_personalize_with_context(campaign_name: str, profileId: str, context_field: str, timeofday):
    print(f"n== {context_field}:{timeofday} against {campaign_name} for {profileId} ===")
    arn = f"arn:aws:personalize:{REGION}:{ACCOUNT}:campaign/{campaign_name}"

    kwargs = {"campaignArn":arn, "userId":profileId}
    if timeofday is not None:
        kwargs["context"] = { context_field: timeofday }

    response = personalize_client.get_recommendations(**kwargs)

    for item in response['itemList']:
        id = item['itemId']
        print(f"{id}")

...

call_personalize_with_context("test_campaign", "john123", "TIME_OF_DAY", "morning")
call_personalize_with_context("test_campaign", "john123", "TIME_OF_DAY", "afternoon")

With the old recipe there is some variation in the result based on the context. However, if I train a model with the same data using the aws-user-personalization-v2 recipe it results in absolutely no variation.

Is there any differences in the way that I should use the new recipe with context? Should the schema be defined differently?

Any assistance would be appreciated.