Reproduce “Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval” on OpenResearch

About this paper

Haolun Wu, Ofer Meshi, Masrour Zoghi, Fernando Diaz, Xue Liu, Craig Boutilier, Maryam Karimzadehgan

A new density-based user representation using Gaussian Process Regression captures diverse and dynamic user interests for personalized retrieval. Developed by researchers at Google Research, McGill University, and Mila, this approach consistently outperforms point-based methods in multi-interest coverage and relevance metrics across three datasets, while also leveraging uncertainty for improved online exploration.

Key insights

• Representing user interests as a continuous density function (DUR) rather than discrete points (SUR/MUR) inherently captures multi-faceted preferences and adapts without manual tuning.
• Gaussian Process Regression naturally provides uncertainty estimates, enabling principled exploration-exploitation trade-offs in recommendations.
• Pre-training semantically rich item embeddings, particularly with an auxiliary item-category association loss, is crucial for effective multi-interest retrieval and efficient GPR application.

Summary by alphaXiv.