BLUE framework bridges user profiles and embeddings for better recommendations

Personalized systems rely on effective user representations, but existing methods face a trade-off: latent embeddings are powerful for retrieval but uninterpretable, while textual profiles are interpretable but hard to optimize. A new paper from researchers (Zhaoxuan Tan, Xiang Zhai, Yan Zhu, Meng Jiang, Mohamed Hammad) introduces BLUE (Bridging texuaL profiles and latent User Embeddings), a reinforcement learning framework that bridges this gap. BLUE uses a profiler LLM to generate textual user profiles from interaction history, while an embedding model provides reward signals to push profiles closer to positive items and farther from negative ones in embedding space. A text-space supervision signal based on next-item prediction ensures profiles remain semantically meaningful.

Experiments on Amazon Reviews 2023 and Google Local Reviews show BLUE consistently outperforms strong baselines in zero-shot sequential recommendation settings, under both frozen and trainable embedding conditions. Notably, BLUE achieves clear gains in cross-domain transfer, demonstrating generalization ability of learned profiles. The generated profiles also provide better context for question answering compared to raw histories or alternative profile methods. This work offers a practical path to unify interpretability and discriminative power in user representation, with implications for recommendation systems, search personalization, and AI assistants.