I'm building a recommendation system using a neural network, and I'm struggling with how to handle sparse user-item interaction data efficiently. What are the best practices for preprocessing and modeling this?

I'm working on a project for an e-commerce site where I need to recommend products based on user purchase history. The interaction matrix is very sparse (most users have only a few interactions), and I'm using collaborative filtering with a neural network in PyTorch. I've tried using embedding layers, but the model isn't learning well, and training is slow. I've already normalized the data and split it into train/val sets. What should I focus on to improve performance?