Performing targeted searching based on a user profile
Topics: Microsoft, News and Discover, Personalization, Probably in use, Search Intent, User Signals
This Microsoft patent describes a system and method for creating a highly targeted search experience driven by deep user understanding rather than just keywords. It works by generating a user embedding (a numerical representation of a user’s interests) and using it to quickly compare against document embeddings across a vast corpus. This process efficiently narrows down a massive number of documents to only the most semantically relevant subset, which is then passed to a high-quality ranking model, such as a transformer model, to recommend niche, timely, and highly personalized content.
