Indexing algorithm based on storing additional distances in metric space for multi-vantage-point tree

Abstract

Introduction: The similarity search paradigm is used in various computational tasks, such as classification, data mining, pattern recognition, etc. Currently, the technology of tree-like metric access methods occupies a significant place among search algorithms. The classical problem of reducing the time of similarity search in metric space is relevant for modern systems when processing big complex data. Due to multidimensional nature of the search algorithm effectiveness problem, local research in this direction is in demand, constantly bringing useful results. Purpose: To reduce the computational complexity of tree search algorithms in problems involving metric proximity. Results: We developed a search algorithm for a multi-vantage-point tree, based on the priority node-processing queue. We mathematically formalized the problems of additional calculations and ways to solve them. To improve the performance of similarity search, we have proposed procedures for forming a priority queue of processing nodes and reducing the number of intersections of same level nodes. Structural changes in the multi-vantage-point tree and the use of minimum distances between vantage points and node subtrees provide better search efficiency. More accurate determination of the distance from the search object to the nodes and the fact that the search area intersects with a tree node allows you to reduce the amount of calculations. Practical relevance: The resulting search algorithms need less time to process information due to an insignificant increase in memory requirements. Reducing the information processing time expands the application boundaries of tree metric indexing methods in search problems involving large data sets.