In elbamos/largeVis: High-Quality Visualizations of Large, High-Dimensional Datasets

This is an implementation of the largeVis algorithm described in (https://arxiv.org/abs/1602.00370). It also incorporates:

• A very fast algorithm for estimating k-nearest neighbors, implemented in C++ with Rcpp and OpenMP. See the Benchmarks file for performance details.
• Efficient implementations of the clustering algorithms:
  • HDBSCAN
  • OPTICS
  • DBSCAN
• Functions for visualizing manifolds like this.

News Highlights

• Version 0.1.10 re-adds clustering, and also adds momentum training to largeVis, as well as a host of other features and improvements.
• Version 0.1.9.1 has been accepted by CRAN. Much grattitude to Uwe Ligges and Kurt Hornik for their assistance, advice, and patience.

Some Examples

Clustering With HDBSCAN

library(clusteringdatasets)
library(largeVis, quietly = TRUE)
library(ggplot2)
data(Aggregation)
vis <- largeVis(t(as.matrix(Aggregation[, 1:2])), sgd_batches = 1)
clusters <- hdbscan(vis, K = 2)

theme_set(
  theme_bw() %+replace%
  theme(
    legend.title = element_text(size = rel(0.8),
                              face = "bold"),
    legend.margin = unit(0, "cm"),
    legend.position = "right",
    legend.key.size = unit(2, "lines"),
    legend.text = element_text(size = unit(8, "points")), 
    axis.title.y = element_text(angle = 90),
    axis.text = element_text(size = rel(0.7)),
    plot.margin = unit(c(0, 0.5, 1, 0), "lines"), 
    axis.title = element_text(size = rel(0.8),
                              face = "bold"),
    title = element_text(size = rel(0.9))
  ) 
)

show(gplot(clusters, Aggregation[, 1:2]) +
            scale_x_continuous("", labels = NULL) +
            scale_y_continuous("", labels = NULL))

Visualize Embeddings

Building Notes

• Note on R 3.4: Before R 3.4, the CRAN binaries were likely to have been compiled without OpenMP, and getting OpenMP to work on Mac OS X was somewhat tricky. This should all have changed (for the better) with R 3.4, which natively using clang 4.0 by default. Since R 3.4 is new, I'm not able to provide advice, but am interested in hearing of any issues and any workarounds to issues that you may discover.

elbamos/largeVis documentation built on May 16, 2019, 2:58 a.m.