This repository tests the performance of the kmknn package for detecting nearest neighbours. Speed comparisons are performed to existing R packages, mostly based on the ANN library (i.e., RANN and FNN). We focus on performance for data sets with moderately high (10-50) dimensions, as discussed by Wang (2012).
Scenarios in simulations/ include:
sim_hypercube.R, consisting of uniformly distributed points in a hypercube.sim_gaussclust.R, consisting of Gaussian clusters.sim_helical.R, consisting of a helical trajectory.
Some of the scripts have tunable parameters that should be specified by the calling process.
This is controlled during job submission, which can be executed by calling submitter.sh for SLURM clusters.
Each dataset in real/ should contain:
proc_*.R, which processes the data into a RDS file for nearest neighbor detection.run_*.R, which runs the algorithm timings on the processed data.
Current data sets are:
- PBMC 68K single-cell RNA-seq data from 10X Genomics
- MNIST data sets of handwritten digits
The plot_results.R scripts in both directories will generate summary plots for each scenario/dataset.
Each plot will show the effect of dimensionality and choice of k, as well as the number of points for the simulations.