Apply hard business rules and an exploration layer. Don’t let raw math ruin the UX. Implement immediate post-processing filters to block out-of-stock, irrelevant, or repetitive categories. Additionally, inject a 10% mix of globally trending or popular items to break bad feedback loops and help the model safely discover new user interests.

Upgrade to a Graph Neural Network (GNN) and tune your weights. Standard matrix factorization misses complex, multi-hop user-item relationships. Transitioning to a GNN (like LightGCN) will better capture deep structural similarities. Also, ensure your loss function downweights superficial clicks and heavily rewards high-value interactions like purchases or long dwell times.

Switch to a Hybrid Retrieval/Ranking system. Pure collaborative filtering fails at scale due to data sparsity. Use it only for the initial retrieval step to get a broad pool of items. Then, pass those items through a second ranking layer (like XGBoost or a Deep & Cross Network) that filters and scores them based on real-time metadata, user context, and categories.