Amazon Personalize Data Requirements

I recently helped a customer in a very similar situation. They had a small set of interactions, but not nearly enough for Personalize. When using Personalize, it was clear that Personalize resorted to its Popularity Count model (which happens when the other recipes can't provide good recommendations).

The customer did have metadata available, which wasn't great quality, but at least provided a bigger dataset than the interactions. I built a prototype with their small interaction dataset and their larger metadata dataset using the LightFM package, which is a variation of factorization machines. The results were promising, despite the bad data quality.

In the case of this customer, their business model meant they would never be able to gather enough interaction data for Personalize, so they are exploring the option of expanding upon the prototype I built. However, if your customer would be able to gather sufficient interaction data over time, then it is likely that Personalize would eventually provide better results. It could be worthwhile implementing an interim solution which places more emphasis on metadata, while gathering data to eventually move to Personalize.