Preference Discerning in Recommender Systems: Generative Retrieval for Personalized Recommendations
The episode delves into Mender, a generative retrieval model designed to enhance recommendation systems by integrating user preferences expressed in natural language. Mender employs a two-phase approach, known as preference approximation and preference conditioning, enabling highly personalized recommendations. This system effectively handles both positive and negative preferences, achieving a significant improvement in accuracy metrics—up to 45%—compared to traditional models. The model leverages semantic IDs to represent items and has been evaluated on various benchmarks, showcasing superior performance in areas such as fine-grained and coarse-grained steering, sentiment adherence, and history consolidation. Its application in e-commerce results in advanced personalization and increased conversions.