Random projection trees have proven to be effective for approximate nearest neighbor searches in high dimensional spaces where conventional methods are not applicable due to excessive usage of memory and computational time. We show that building multiple trees on the same data can improve the performance even further, without significantly increasing the total computational cost of queries when executed in a modern parallel computing environment. Our experiments identify suitable parameter values to achieve accurate searches with extremely fast query times, while also retaining a feasible complexity for index construction.
View on arXiv