Norm Adjusted Proximity Graph for Fast Inner Product Retrieval

Abstract

Efficient inner product search on embedding vectors is often the vital stage for online ranking services, such as recommendation and information retrieval. Recommendation algorithms, e.g., matrix factorization, typically produce latent vectors to represent users or items. The recommendation services are conducted by retrieving the most relevant item vectors given the user vector, where the relevance is often defined by inner product. Therefore, developing efficient recommender systems often requires solving the so-called maximum inner product search (MIPS) problem. In the past decade, there have been many studies on efficient MIPS algorithms. This task is challenging in part because the inner product does not follow the triangle inequality of metric space.