Performance Analysis of Graph-Based Methods for Exact and Approximate Similarity Search in Metric Spaces

Abstract

Similarity searches are widely used to retrieve complex data, such as images, videos, and georeferenced data. Recently, graph-based methods have emerged as a very efficient alternative for similarity retrieval. However, to the best of our knowledge, there are no previous works with experimental analysis on a comprehensive number of graph-based methods using the same search algorithm and execution environment. In this work, we survey the main graph-based types currently employed for similarity searches and present an experimental evaluation of the most representative graphs on a common platform. We evaluate the relative performance behavior of the tested graph-based methods with respect to the main construction and query parameters for a variety of real-world datasets. Our experimental results provide a quantitative view of the exact search compared to accurate setups for approximate search. These results reinforce the tradeoff between graph construction cost and search performance according to the construction and search parameters. With respect to the approximate methods, the Navigable Small World graph (NSW) presented the highest recall rates. Nevertheless, given a recall rate, there is no winner graph for query performance.