High Fidelity Textual User Representation over Heterogeneous Sources via Reinforcement Learning

Researchers at LinkedIn developed a new AI method using reinforcement learning to create unified, concise text profiles for users by combining data from different sources like job profiles, activity logs, and searches. The system learns from user clicks and applications as rewards, improving key business metrics. This provides a scalable, label-free solution for building interpretable user representations that work directly with modern large language model-based recommendation systems.