negative sampling from small negative observations for recommendation system

Follow the full discussion on Reddit.
I am working on a recommendation system on a user item interaction matrix based on implicit feedback (binary), and I have positive observed interactions and a very small amount of negative observed interactions. For both user and item there are vector features available as embeddings. Are there any methods that I can perform negative sampling through the negative observations? I have heard of methods like contrastive learning but not aware of ways to integrate existing negative observations.