xiaozhxiong / gbnns_theory

Code for ICML2020 paper: ''Graph-based Nearest Neighbor Search: From Practice to Theory''

Graph-based Nearest Neighbor Search: From Practice to Theory

Code for reproducing synthetic experiments from ICML2020 paper

Files description

major_test.cpp is code for reproducing the main illustration of how analyzed heuristics and proposed methods work with synthetic datasets.

Code from links_type.cpp analyze the effect of kNN and KL approximation and number_links_test.cpp illustrate the effect of the number of long-range edges.

For reproducing distribution of the distance to the nearest neighbor use draw_1nn_distr.cpp

You can find results in results folder, with draw_results.ipynb for transformation '.txt' files to pictures.

Data prepare

All programs will build corresponding graphs and datasets if necessary.

What about running?

Most of the functions are simple and straightforward, therefore machine with more CPU (only supported now) preferred.

To run it you need to specify paths to prospective data location in corresponding file and do g++ -Ofast -std=c++11 -fopenmp -march=native -fpic -w -ftree-vectorize file_name_test.cpp -o some_test.exe and ./some_test.exe 16