In a single survey post, Eugene Yan walks through five families of personalization techniques used in recommendation and search systems: contextual bandits, embedding-plus-MLP models, sequence-based models, graph-based approaches, and user embeddings. The post is structured as a whirlwind tour rather than a deep dive, giving practitioners a map of the options available.
Each pattern comes with its own trade-offs. Bandits are lightweight and good for exploration, but they offer only shallow personalization. Embedding-plus-MLP and sequence models are more expressive and can capture complex user-item interactions, at the cost of greater engineering complexity. Graph methods and user embeddings go further by modeling relationships between users and items, or by representing users directly as vectors for retrieval and ranking.
Because there is only one source, there are no conflicting viewpoints to reconcile. Yan's framing is practical: the right pattern depends on the stage of the system, the data available, and the engineering resources at hand. The post serves as a useful starting point for teams deciding which personalization approach to adopt or evolve.