Integrating Textual Profiles with Latent User Representations for Enhanced Personalization

Personalized recommendation systems depend heavily on user representations that link past behaviors to future suggestions. Traditional approaches often rely on either supervised latent embeddings, which excel in retrieval but lack interpretability, or textual profiles that are easy to understand but difficult to optimize due to limited supervision. To address this issue, we propose BLUE, a reinforcement learning framework that harmonizes these two user representation types. By utilizing a Large Language Model (LLM), BLUE generates textual profiles based on user interaction history, while an embedding model supplies reward signals. This alignment encourages the textual representations to cluster around positive items and distance from negative ones within the embedding space. Additionally, we introduce a supervision mechanism based on next-item prediction to ensure that the generated profiles are both meaningful and effective for retrieval tasks. Evaluations on datasets like Amazon Reviews 2023 and Google Local Reviews in zero-shot settings reveal that BLUE consistently surpasses established baselines, demonstrating its robustness in cross-domain scenarios. The profiles produced by BLUE also offer enhanced personalized contexts for question answering compared to traditional methods. Overall, BLUE effectively combines the interpretability of textual profiles with the discriminative power of latent embeddings for improved personalization.