Theoretical Analysis of Two Tower Recommendation Models

A new paper on arXiv (2403.00802v2) provides a theoretical analysis of two tower recommendation models, which are widely used in production-grade recommender systems by online media services such as Netflix, Pinterest, and Amazon. These systems learn embeddings of users and items in a low-dimensional space using two deep neural networks, enabling predictions of user feedback. Despite their popularity, the theoretical behaviors of these models have been largely unexplored. The paper studies the asymptotic behaviors of two tower models in two-stage recommenders, demonstrating strong convergence to the optimal recommender system. It establishes theoretical properties and statistical assurances for the two tower recommender, and shows that faster convergence is achieved by relying on the intrinsic dimensions of the data. The findings contribute to a deeper understanding of these models, potentially guiding future improvements in recommendation systems.