Learned Metric Index - Proposition of learned indexing for unstructured data

摘要

The main paradigm of similarity searching in metric spaces has remained mostly unchanged for decades - data objects are organized into a hierarchical structure according to their mutual distances, using representative pivots to reduce the number of distance computations needed to efficiently search the data. We propose an alternative to this paradigm, using machine learning models to replace pivots, thus posing similarity search as a classification problem, which stands in for numerous expensive distance computations. Even a relatively naive implementation of this idea is more than competitive with state-of-the-art methods in terms of speed and recall, proving the concept as viable and showing great potential for its future development. (C) 2021 Elsevier Ltd. All rights reserved.