Content-based recommendation using Amazon Personalize

Question

I have no user in my platform still I foresee investing time in building a rec-sys based on Amazon Personalize (AP) is worth doing in a long run.

Seems, AP leverage meta-data so I assume the underlying algorithm is not a vanilla Collaborative Filtering (NNF) and rather a hybrid approach.

However, in the git repo here . They list following as indicator of a bad-fit:

1. if one only has explicit meta data (the case for content-based recommendation) AP is not a good fit.

2. low data volumes for user is neither a good fit. So I wonder, if it is a hybrid approach (combining both CF and content-based) why they don't recommend using AP here ?

Here is my question: Is there any way to get content-based recommendation out of AP at least for first 100 users and then move towards CF / hybrdi ?

Answer 1

All item-based recommendations recipes in Amazon Personalize currently require interaction data. Item metadata and user metadata are optional and whether they are used in training or not varies based on the recipe. For example, the user-personalization recipe will use all three datasets if provided while the trending now recipe only trains in the interactions dataset. Therefore, you can say that Personalize uses collaborative filtering but is also capable of using a hybrid approach depending on the recipe and the data provided. However, Personalize does not provide a purely content filtering based algorithm/recipe.

The reason for the low data (volume) call out in the Personalize cheat sheet as being a potentially bad fit is because low levels of interaction data typically do not produce good models. In other words, if the application or use case just doesn't involve a reasonable amount of user interaction with recommended items, there won't be a lot of signal to go on to predict intent/interest.