LightFM

A hybrid recommendation algorithm that bridges collaborative filtering and content-based models using factorized representations.

Developed by Lyst, LightFM solves the cold-start problem by representing users and items as linear combinations of their metadata features. It outperforms standard matrix factorization (MF) by learning embeddings through a weighted approximate-rank pairwise (WARP) loss function. This architecture allows the model to generalize to unseen items (new inventory) and users (new sign-ups) while maintaining the high-performance latent representation benefits of traditional CF. It is a go-to choice for production systems requiring fast, Python-based implementations that handle sparse interaction data alongside rich feature sets.

https://github.com/lyst/lightfm