damirpolat / dbscan

Density Based Clustering of Applications with Noise (DBSCAN) and Related Algorithms - R package

dbscan - Density Based Clustering of Applications with Noise (DBSCAN) and Related Algorithms - R package

This R package provides a fast C++ (re)implementation of several density-based algorithms with a focus on the DBSCAN family for clustering spatial data. The package includes:

Clustering

• DBSCAN: Density-based spatial clustering of applications with noise.
• HDBSCAN: Hierarchical DBSCAN with simplified hierarchy extraction.
• OPTICS/OPTICSXi: Ordering points to identify the clustering structure clustering algorithms.
• FOSC: Framework for Optimal Selection of Clusters for unsupervised and semisupervised clustering of hierarchical cluster tree.
• Jarvis-Patrick clustering
• SNN Clustering: Shared Nearest Neighbor Clustering.

Outlier Detection

• LOF: Local outlier factor algorithm.
• GLOSH: Global-Local Outlier Score from Hierarchies algorithm.

Fast Nearest-Neighbor Search (using kd-trees)

• kNN search
• Fixed-radius NN search

The implementations use the kd-tree data structure (from library ANN) for faster k-nearest neighbor search, and are typically faster than the native R implementations (e.g., dbscan in package fpc), or the implementations in WEKA, ELKI and Python's scikit-learn.

Installation

Stable CRAN version: install from within R with

install.packages("dbscan")

Current development version: Download package from AppVeyor or install from GitHub (needs devtools).

library("devtools")
install_github("mhahsler/dbscan")

Usage

Load the package and use the numeric variables in the iris dataset

library("dbscan")

data("iris")
x <- as.matrix(iris[, 1:4])

Run DBSCAN

db <- dbscan(x, eps = .4, minPts = 4)
db

DBSCAN clustering for 150 objects.
Parameters: eps = 0.4, minPts = 4
The clustering contains 4 cluster(s) and 25 noise points.

 0  1  2  3  4 
25 47 38 36  4 

Available fields: cluster, eps, minPts

Visualize results (noise is shown in black)

pairs(x, col = db$cluster + 1L)

Calculate LOF (local outlier factor) and visualize (larger bubbles in the visualization have a larger LOF)

lof <- lof(x, k = 4)
pairs(x, cex = lof)

Run OPTICS

opt <- optics(x, eps = 1, minPts = 4)
opt

OPTICS clustering for 150 objects.
Parameters: minPts = 4, eps = 1, eps_cl = NA, xi = NA
Available fields: order, reachdist, coredist, predecessor, minPts, eps, eps_cl, xi

Extract DBSCAN-like clustering from OPTICS and create a reachability plot (extracted DBSCAN clusters at eps_cl=.4 are colored)

opt <- extractDBSCAN(opt, eps_cl = .4)
plot(opt)

Extract a hierarchical clustering using the Xi method (captures clusters of varying density)

opt <- extractXi(opt, xi = .05)
opt
plot(opt)

Run HDBSCAN (captures stable clusters)

hdb <- hdbscan(x, minPts = 4)
hdb

HDBSCAN clustering for 150 objects.
Parameters: minPts = 4
The clustering contains 2 cluster(s) and 0 noise points.

  1   2 
100  50 

Available fields: cluster, minPts, cluster_scores, membership_prob, outlier_scores, hc

Visualize the results as a simplified tree

plot(hdb, show_flat = T)

See how well each point corresponds to the clusters found by the model used

  colors <- mapply(function(col, i) adjustcolor(col, alpha.f = hdb$membership_prob[i]), 
                   palette()[hdb$cluster+1], seq_along(hdb$cluster))
  plot(x, col=colors, pch=20)

License

The dbscan package is licensed under the GNU General Public License (GPL) Version 3. The OPTICSXi R implementation was directly ported from the ELKI framework's Java implementation (GNU AGPLv3), with explicit permission granted by the original author, Erich Schubert.

Further Information

• List of changes from NEWS.md