kmeans-clusterer 0.4.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/lib/kmeans-clusterer.rb +282 -0
- metadata +44 -0
checksums.yaml
ADDED
@@ -0,0 +1,7 @@
|
|
1
|
+
---
|
2
|
+
SHA1:
|
3
|
+
metadata.gz: daff89d4293b6131cf1c684c9a5ef3a3d9dfa878
|
4
|
+
data.tar.gz: 7917de7e84065b9c55c4c501bb2587f6e86aff7b
|
5
|
+
SHA512:
|
6
|
+
metadata.gz: 795ac3f9fb65b6a40adbb67264413fc3819bf828de13edef857ef340d5e2467de4807e35fb87277bd6aeec687ab93f11b35c0b34a92922d0b5c609a0e731be74
|
7
|
+
data.tar.gz: ce6dc360936e0a3fcdc4be3b4c3fd207d0b0df67ed45c36ea2b476e473eedd11b7fea35c060eabca38f733325d1aba6fa30defa56601bdcac6c65d68c27d8145
|
@@ -0,0 +1,282 @@
|
|
1
|
+
require 'narray'
|
2
|
+
|
3
|
+
class KMeansClusterer
|
4
|
+
|
5
|
+
# Euclidean distance function. Requires instances of NArray as args
|
6
|
+
Distance = -> (a, b) { NMath.sqrt ((a - b)**2).sum(0) }
|
7
|
+
CalculateCentroid = -> (a) { a.mean(1) }
|
8
|
+
|
9
|
+
class Point
|
10
|
+
attr_reader :data
|
11
|
+
attr_accessor :cluster, :label
|
12
|
+
|
13
|
+
def initialize data, label = nil
|
14
|
+
@data = NArray.to_na data
|
15
|
+
@label = label
|
16
|
+
end
|
17
|
+
|
18
|
+
def [] index
|
19
|
+
@data[index]
|
20
|
+
end
|
21
|
+
|
22
|
+
def to_a
|
23
|
+
@data.to_a
|
24
|
+
end
|
25
|
+
|
26
|
+
def to_s
|
27
|
+
to_a.to_s
|
28
|
+
end
|
29
|
+
|
30
|
+
def dimension
|
31
|
+
@data.length
|
32
|
+
end
|
33
|
+
end
|
34
|
+
|
35
|
+
|
36
|
+
class Cluster
|
37
|
+
attr_reader :centroid, :points
|
38
|
+
attr_accessor :label
|
39
|
+
|
40
|
+
def initialize centroid, label = nil
|
41
|
+
@centroid = centroid
|
42
|
+
@label = label
|
43
|
+
@points = []
|
44
|
+
end
|
45
|
+
|
46
|
+
def recenter
|
47
|
+
if @points.empty?
|
48
|
+
0
|
49
|
+
else
|
50
|
+
old_centroid = @centroid
|
51
|
+
@centroid = calculate_centroid_from_points
|
52
|
+
Distance.call @centroid.data, old_centroid.data
|
53
|
+
end
|
54
|
+
end
|
55
|
+
|
56
|
+
def << point
|
57
|
+
point.cluster = self
|
58
|
+
@points << point
|
59
|
+
end
|
60
|
+
|
61
|
+
def reset_points
|
62
|
+
@points = []
|
63
|
+
end
|
64
|
+
|
65
|
+
def sorted_points
|
66
|
+
distances = Distance.call points_narray, centroid.data
|
67
|
+
@points.sort_by.with_index {|c, i| distances[i] }
|
68
|
+
end
|
69
|
+
|
70
|
+
def sum_of_squares_error
|
71
|
+
if @points.empty?
|
72
|
+
0
|
73
|
+
else
|
74
|
+
distances = Distance.call points_narray, centroid.data
|
75
|
+
(distances**2).sum
|
76
|
+
end
|
77
|
+
end
|
78
|
+
|
79
|
+
def sum_of_distances
|
80
|
+
return 0 if @points.empty?
|
81
|
+
Distance.call(points_narray, centroid.data).sum
|
82
|
+
end
|
83
|
+
|
84
|
+
def dissimilarity point
|
85
|
+
distances = Distance.call points_narray, point.data
|
86
|
+
distances.sum / distances.length.to_f
|
87
|
+
end
|
88
|
+
|
89
|
+
private
|
90
|
+
def calculate_centroid_from_points
|
91
|
+
data = CalculateCentroid.call points_narray
|
92
|
+
Point.new data
|
93
|
+
end
|
94
|
+
|
95
|
+
def points_narray
|
96
|
+
NArray.to_na @points.map(&:data)
|
97
|
+
end
|
98
|
+
end
|
99
|
+
|
100
|
+
|
101
|
+
def self.run k, data, opts = {}
|
102
|
+
raise(ArgumentError, "k cannot be greater than the number of points") if k > data.length
|
103
|
+
|
104
|
+
data = if opts[:scale_data]
|
105
|
+
scale_data data
|
106
|
+
else
|
107
|
+
data.map {|row| NArray.to_na(row).to_f}
|
108
|
+
end
|
109
|
+
|
110
|
+
runcount = opts[:runs] || 10
|
111
|
+
errors = []
|
112
|
+
|
113
|
+
runs = runcount.times.map do |i|
|
114
|
+
km = new(k, data, opts).run
|
115
|
+
error = km.error
|
116
|
+
if opts[:log]
|
117
|
+
puts "[#{i + 1}] #{km.iterations} iter\t#{km.runtime.round(2)}s\t#{error.round(2)} err"
|
118
|
+
end
|
119
|
+
errors << error
|
120
|
+
km
|
121
|
+
end
|
122
|
+
|
123
|
+
runs.sort_by.with_index {|run, i| errors[i] }.first
|
124
|
+
end
|
125
|
+
|
126
|
+
# see scikit-learn scale and _mean_and_std methods
|
127
|
+
def self.scale_data data
|
128
|
+
nadata = NArray.to_na(data).to_f
|
129
|
+
mean = nadata.mean(1)
|
130
|
+
std = nadata.rmsdev(1)
|
131
|
+
std[std.eq(0)] = 1.0 # so we don't divide by 0
|
132
|
+
nadata = (nadata - mean) / std
|
133
|
+
# convert back to an array, containing NArrays for each row
|
134
|
+
data.length.times.map {|i| nadata[true, i] }
|
135
|
+
end
|
136
|
+
|
137
|
+
|
138
|
+
attr_reader :k, :points, :clusters, :iterations, :runtime
|
139
|
+
|
140
|
+
|
141
|
+
def initialize k, data, opts = {}
|
142
|
+
@k = k
|
143
|
+
@init = opts[:init] || :kmpp
|
144
|
+
labels = opts[:labels] || []
|
145
|
+
|
146
|
+
@points = data.map.with_index do |instance, i|
|
147
|
+
Point.new instance, labels[i]
|
148
|
+
end
|
149
|
+
|
150
|
+
init_clusters
|
151
|
+
end
|
152
|
+
|
153
|
+
def run
|
154
|
+
start_time = Time.now
|
155
|
+
@iterations, @runtime = 0, 0
|
156
|
+
|
157
|
+
loop do
|
158
|
+
@iterations +=1
|
159
|
+
|
160
|
+
centroids = get_cluster_centroids
|
161
|
+
|
162
|
+
@points.each do |point|
|
163
|
+
distances = Distance.call(centroids, point.data)
|
164
|
+
cluster = @clusters.sort_by.with_index {|c, i| distances[i] }.first
|
165
|
+
cluster << point
|
166
|
+
end
|
167
|
+
|
168
|
+
moves = clusters.map(&:recenter)
|
169
|
+
|
170
|
+
break if moves.max < 0.001 # i.e., no movement
|
171
|
+
break if @iterations >= 300
|
172
|
+
|
173
|
+
clusters.each(&:reset_points)
|
174
|
+
end
|
175
|
+
|
176
|
+
@runtime = Time.now - start_time
|
177
|
+
self
|
178
|
+
end
|
179
|
+
|
180
|
+
def error
|
181
|
+
@clusters.map(&:sum_of_squares_error).reduce(:+)
|
182
|
+
end
|
183
|
+
|
184
|
+
def closest_cluster point = origin
|
185
|
+
sorted_clusters(point).first
|
186
|
+
end
|
187
|
+
|
188
|
+
def sorted_clusters point = origin
|
189
|
+
point = Point.new(point) unless point.is_a?(Point)
|
190
|
+
centroids = get_cluster_centroids
|
191
|
+
distances = Distance.call(centroids, point.data)
|
192
|
+
@clusters.sort_by.with_index {|c, i| distances[i] }
|
193
|
+
end
|
194
|
+
|
195
|
+
def origin
|
196
|
+
Point.new Array.new(@points[0].dimension, 0)
|
197
|
+
end
|
198
|
+
|
199
|
+
def silhouette_score
|
200
|
+
return 1.0 if @clusters.length < 2
|
201
|
+
|
202
|
+
scores = @points.map do |point|
|
203
|
+
acluster, bcluster = sorted_clusters(point).slice(0,2)
|
204
|
+
a = acluster.dissimilarity(point)
|
205
|
+
b = bcluster.dissimilarity(point)
|
206
|
+
(b - a) / [a,b].max
|
207
|
+
end
|
208
|
+
|
209
|
+
scores.reduce(:+) / scores.length # mean score for all points
|
210
|
+
end
|
211
|
+
|
212
|
+
private
|
213
|
+
def init_clusters
|
214
|
+
case @init
|
215
|
+
when :random
|
216
|
+
random_cluster_init
|
217
|
+
when Array
|
218
|
+
custom_cluster_init
|
219
|
+
else
|
220
|
+
kmpp_cluster_init
|
221
|
+
end
|
222
|
+
end
|
223
|
+
|
224
|
+
# k-means++
|
225
|
+
def kmpp_cluster_init
|
226
|
+
@clusters = []
|
227
|
+
pick = rand(@points.length)
|
228
|
+
centroid = Point.new @points[pick].data.to_a
|
229
|
+
@clusters << Cluster.new(centroid, 1)
|
230
|
+
|
231
|
+
while @clusters.length < @k
|
232
|
+
centroids = get_cluster_centroids
|
233
|
+
|
234
|
+
d2 = @points.map do |point|
|
235
|
+
dists = Distance.call centroids, point.data
|
236
|
+
dists.min**2 # closest cluster distance, squared
|
237
|
+
end
|
238
|
+
|
239
|
+
d2 = NArray.to_na d2
|
240
|
+
probs = d2 / d2.sum
|
241
|
+
cumprobs = probs.cumsum
|
242
|
+
r = rand
|
243
|
+
# pick = cumprobs.to_a.index {|prob| r < prob }
|
244
|
+
pick = (cumprobs >= r).where[0]
|
245
|
+
centroid = Point.new @points[pick].data.to_a
|
246
|
+
cluster = Cluster.new(centroid, @clusters.length + 1)
|
247
|
+
@clusters << cluster
|
248
|
+
end
|
249
|
+
end
|
250
|
+
|
251
|
+
def custom_cluster_init
|
252
|
+
@clusters = @init.map.with_index do |instance, i|
|
253
|
+
point = Point.new NArray.to_na(instance).to_f
|
254
|
+
Cluster.new point, i+1
|
255
|
+
end
|
256
|
+
end
|
257
|
+
|
258
|
+
def random_cluster_init
|
259
|
+
@clusters = pick_k_random_points.map.with_index {|centroid, i| Cluster.new centroid, i+1 }
|
260
|
+
end
|
261
|
+
|
262
|
+
def pick_k_random_points
|
263
|
+
pick_k_random_indexes.map {|i| Point.new @points[i].data.to_a }
|
264
|
+
end
|
265
|
+
|
266
|
+
def pick_k_random_indexes
|
267
|
+
@points.length.times.to_a.shuffle.slice(0, @k)
|
268
|
+
end
|
269
|
+
|
270
|
+
def get_cluster_centroids
|
271
|
+
NArray.to_na @clusters.map {|c| c.centroid.data }
|
272
|
+
end
|
273
|
+
end
|
274
|
+
|
275
|
+
class KMediansClusterer < KMeansClusterer
|
276
|
+
Distance = -> (a, b) { (a - b).abs.sum(0) }
|
277
|
+
CalculateCentroid = -> (a) { a.rot90.median(0) }
|
278
|
+
|
279
|
+
def error
|
280
|
+
@clusters.map(&:sum_of_distances).reduce(:+)
|
281
|
+
end
|
282
|
+
end
|
metadata
ADDED
@@ -0,0 +1,44 @@
|
|
1
|
+
--- !ruby/object:Gem::Specification
|
2
|
+
name: kmeans-clusterer
|
3
|
+
version: !ruby/object:Gem::Version
|
4
|
+
version: 0.4.0
|
5
|
+
platform: ruby
|
6
|
+
authors:
|
7
|
+
- Geoff Buesing
|
8
|
+
autorequire:
|
9
|
+
bindir: bin
|
10
|
+
cert_chain: []
|
11
|
+
date: 2015-01-29 00:00:00.000000000 Z
|
12
|
+
dependencies: []
|
13
|
+
description: k-means/k-medians clustering. Uses NArray for fast calculations.
|
14
|
+
email: gbuesing@gmail.com
|
15
|
+
executables: []
|
16
|
+
extensions: []
|
17
|
+
extra_rdoc_files: []
|
18
|
+
files:
|
19
|
+
- lib/kmeans-clusterer.rb
|
20
|
+
homepage: https://github.com/gbuesing/kmeans-clusterer
|
21
|
+
licenses:
|
22
|
+
- MIT
|
23
|
+
metadata: {}
|
24
|
+
post_install_message:
|
25
|
+
rdoc_options: []
|
26
|
+
require_paths:
|
27
|
+
- lib
|
28
|
+
required_ruby_version: !ruby/object:Gem::Requirement
|
29
|
+
requirements:
|
30
|
+
- - ">="
|
31
|
+
- !ruby/object:Gem::Version
|
32
|
+
version: '0'
|
33
|
+
required_rubygems_version: !ruby/object:Gem::Requirement
|
34
|
+
requirements:
|
35
|
+
- - ">="
|
36
|
+
- !ruby/object:Gem::Version
|
37
|
+
version: '0'
|
38
|
+
requirements: []
|
39
|
+
rubyforge_project:
|
40
|
+
rubygems_version: 2.4.5
|
41
|
+
signing_key:
|
42
|
+
specification_version: 4
|
43
|
+
summary: k-means/k-medians clustering
|
44
|
+
test_files: []
|