Patterns for Personalization in Recommendations and Search
Eugene Yan
This article explains patterns for personalization in recommendations and search systems, covering contextual bandits (which continuously learn through exploration and exploitation) and embedding-based deep learning approaches (which map features to vectors then process through neural networks). Netflix uses contextual bandits for image selection with a take-fraction metric of quality plays per impressions, while YouTube applies embedding pooling and mean-pooling strategies across millions of video candidates. These techniques allow systems to personalize experiences by learning user preferences through continuous adaptation or by compressing variable-length user histories into fixed-size vectors for ranking.
Why it matters
A whirlwind tour of bandits, embedding+MLP, sequences, graph, and user embeddings.
