Generating product latent vectors from interactions

Question

I would like to create product embeddings, i.e product vectors that describe the product in terms of qualitative attributes & interactions. For qualitative attributes, I have been able to use NLP techniques to create embeddings from product descriptions. Now I want to create interaction embeddings, i.e latent vectors describing my products based on user interactions. I know I should use collaborative filtering & matrix factorization techniques for this purpose, but I don't know how to extract the latent features generated by such models at the end. Models, like AWS Factorization Machines, return recommendations directly but don't return the latent item vectors upon which these recommendations were made. I would like to extract the latent vectors generated by the model to be able to concat them to my qualitative product embeddings and have comprehensive embeddings describing my products in terms of qualitative attributes & interactions. Any idea what model can I use for the purpose of creating item latent vectors from interactions, or how can I potentially extract latent vectors from such recommendation generation models?