fselector 0.1.2 → 0.2.0
Sign up to get free protection for your applications and to get access to all the features.
- data/LICENSE +1 -1
- data/README.md +14 -12
- data/lib/fselector.rb +11 -10
- data/lib/fselector/{base.rb → algo_base/base.rb} +33 -41
- data/lib/fselector/algo_base/base_CFS.rb +135 -0
- data/lib/fselector/algo_base/base_Relief.rb +130 -0
- data/lib/fselector/algo_base/base_ReliefF.rb +157 -0
- data/lib/fselector/{base_continuous.rb → algo_base/base_continuous.rb} +2 -2
- data/lib/fselector/algo_base/base_discrete.rb +190 -0
- data/lib/fselector/algo_continuous/CFS_c.rb +47 -0
- data/lib/fselector/algo_continuous/ReliefF_c.rb +4 -133
- data/lib/fselector/algo_continuous/Relief_c.rb +3 -103
- data/lib/fselector/algo_discrete/CFS_d.rb +41 -0
- data/lib/fselector/algo_discrete/FastCorrelationBasedFilter.rb +1 -1
- data/lib/fselector/algo_discrete/InformationGain.rb +15 -2
- data/lib/fselector/algo_discrete/ReliefF_d.rb +3 -132
- data/lib/fselector/algo_discrete/Relief_d.rb +3 -103
- data/lib/fselector/entropy.rb +125 -0
- data/lib/fselector/util.rb +22 -2
- metadata +20 -6
- data/lib/fselector/base_discrete.rb +0 -502
@@ -0,0 +1,157 @@
|
|
1
|
+
#
|
2
|
+
# FSelector: a Ruby gem for feature selection and ranking
|
3
|
+
#
|
4
|
+
module FSelector
|
5
|
+
#
|
6
|
+
# base class for extended Relief algorithm (ReliefF), see specialized versions for
|
7
|
+
# discrete feature (ReliefF_d) and continuous feature (ReliefF_c), respectively
|
8
|
+
#
|
9
|
+
# @note applicable to multi-class problem with missing data
|
10
|
+
#
|
11
|
+
# ref: [Estimating Attributes: Analysis and Extensions of RELIEF](http://www.springerlink.com/content/fp23jh2h0426ww45/)
|
12
|
+
#
|
13
|
+
class BaseReliefF < Base
|
14
|
+
#
|
15
|
+
# new()
|
16
|
+
#
|
17
|
+
# @param [Integer] m number of samples to be used
|
18
|
+
# for estimating feature contribution. max can be
|
19
|
+
# the number of training samples
|
20
|
+
# @param [Integer] k number of k-nearest neighbor
|
21
|
+
# @param [Hash] data existing data structure
|
22
|
+
#
|
23
|
+
def initialize(m=nil, k=nil, data=nil)
|
24
|
+
super(data)
|
25
|
+
@m = (m || 30) # default 30
|
26
|
+
@k = (k || 10) # default 10
|
27
|
+
end
|
28
|
+
|
29
|
+
private
|
30
|
+
|
31
|
+
# calculate contribution of each feature (f) across all classes
|
32
|
+
def calc_contribution(f)
|
33
|
+
score = 0.0
|
34
|
+
|
35
|
+
## use all samples if @m not provided
|
36
|
+
#@m = get_sample_size if not @m
|
37
|
+
|
38
|
+
@m.times do
|
39
|
+
# pick a sample at random
|
40
|
+
rs, rk = pick_a_sample_at_random
|
41
|
+
|
42
|
+
# find k nearest neighbor for each class
|
43
|
+
nbrs = find_k_nearest_nb(rs, rk)
|
44
|
+
|
45
|
+
# calc contribution from neighbors
|
46
|
+
score += calc_score(f, rs, rk, nbrs)
|
47
|
+
end
|
48
|
+
|
49
|
+
s = score / @m
|
50
|
+
|
51
|
+
set_feature_score(f, :BEST, s)
|
52
|
+
end # calc_contribution
|
53
|
+
|
54
|
+
|
55
|
+
# pick a sample at random
|
56
|
+
def pick_a_sample_at_random
|
57
|
+
rk = get_classes[rand(get_classes.size)]
|
58
|
+
rks = get_data[rk]
|
59
|
+
|
60
|
+
[ rks[rand(rks.size)], rk ]
|
61
|
+
end # pick_a_sample_at_random
|
62
|
+
|
63
|
+
# # find k nearest neighbors of sample (rs) for each class
|
64
|
+
def find_k_nearest_nb(rs, rk)
|
65
|
+
nbrs = {}
|
66
|
+
|
67
|
+
each_class do |k|
|
68
|
+
res = []
|
69
|
+
|
70
|
+
get_data[k].each do |s|
|
71
|
+
next if s.object_id == rs.object_id # exclude self
|
72
|
+
|
73
|
+
d = diff_sample(rs, s, rk, k)
|
74
|
+
res << [d, s]
|
75
|
+
end
|
76
|
+
|
77
|
+
nbrs[k] = (res.sort { |x, y| x[0] <=> y[0] }[0...@k]).collect { |z| z[1] }
|
78
|
+
end
|
79
|
+
|
80
|
+
nbrs
|
81
|
+
end # find_k_nearest_nb
|
82
|
+
|
83
|
+
|
84
|
+
# difference between two samples
|
85
|
+
def diff_sample(s1, s2, k1, k2)
|
86
|
+
d = 0.0
|
87
|
+
|
88
|
+
each_feature do |f|
|
89
|
+
d += diff_feature(f, s1, s2, k1, k2)**2
|
90
|
+
end
|
91
|
+
|
92
|
+
d
|
93
|
+
end # diff_sample
|
94
|
+
|
95
|
+
|
96
|
+
# difference beween the feature (f) of two samples
|
97
|
+
def diff_feature(f, s1, s2, k1, k2)
|
98
|
+
abort "[#{__FILE__}@#{__LINE__}]: "+
|
99
|
+
"derived ReliefF algo must implement its own diff_feature()"
|
100
|
+
end # diff_feature
|
101
|
+
|
102
|
+
|
103
|
+
# calc probability of missing value (mv)
|
104
|
+
def calc_p(f, mv, k)
|
105
|
+
# cache
|
106
|
+
if not @f2mvp
|
107
|
+
@f2mvp = {}
|
108
|
+
|
109
|
+
each_feature do |f|
|
110
|
+
@f2mvp[f] = {}
|
111
|
+
|
112
|
+
each_class do |k|
|
113
|
+
@f2mvp[f][k] = {}
|
114
|
+
|
115
|
+
fvs = get_feature_values(f).uniq
|
116
|
+
fvs.each do |v|
|
117
|
+
n = 0.0
|
118
|
+
|
119
|
+
get_data[k].each do |s|
|
120
|
+
n += 1 if s.has_key?(f) and s[f] == v
|
121
|
+
end
|
122
|
+
|
123
|
+
@f2mvp[f][k][v] = n/get_data[k].size
|
124
|
+
end
|
125
|
+
end
|
126
|
+
end
|
127
|
+
end
|
128
|
+
|
129
|
+
@f2mvp[f][k][mv]
|
130
|
+
end
|
131
|
+
|
132
|
+
|
133
|
+
# calc feature (f) contribution from neighbors
|
134
|
+
def calc_score(f, rs, rk, nbrs)
|
135
|
+
score = 0.0
|
136
|
+
|
137
|
+
nbrs.each do |k, nbs|
|
138
|
+
if k == rk # near hit
|
139
|
+
nbs.each do |s|
|
140
|
+
score -= (diff_feature(f, rs, s, rk, k)**2/nbs.size)
|
141
|
+
end
|
142
|
+
else # near_miss
|
143
|
+
nbs.each do |s|
|
144
|
+
score += (get_data[k].size/get_sample_size.to_f *
|
145
|
+
diff_feature(f, rs, s, rk, k)**2/nbs.size)
|
146
|
+
end
|
147
|
+
end
|
148
|
+
end
|
149
|
+
|
150
|
+
score
|
151
|
+
end
|
152
|
+
|
153
|
+
|
154
|
+
end # class
|
155
|
+
|
156
|
+
|
157
|
+
end # module
|
@@ -1,5 +1,5 @@
|
|
1
|
-
require File.expand_path(File.dirname(__FILE__) + '
|
2
|
-
require File.expand_path(File.dirname(__FILE__) + '
|
1
|
+
require File.expand_path(File.dirname(__FILE__) + '/../algo_continuous/normalizer.rb')
|
2
|
+
require File.expand_path(File.dirname(__FILE__) + '/../algo_continuous/discretizer.rb')
|
3
3
|
#
|
4
4
|
# FSelector: a Ruby gem for feature selection and ranking
|
5
5
|
#
|
@@ -0,0 +1,190 @@
|
|
1
|
+
#
|
2
|
+
# FSelector: a Ruby gem for feature selection and ranking
|
3
|
+
#
|
4
|
+
module FSelector
|
5
|
+
#
|
6
|
+
# base ranking alogrithm for handling discrete feature
|
7
|
+
#
|
8
|
+
# 2 x 2 contingency table
|
9
|
+
#
|
10
|
+
# c c'
|
11
|
+
# ---------
|
12
|
+
# f | A | B | A+B
|
13
|
+
# |---|---|
|
14
|
+
# f' | C | D | C+D
|
15
|
+
# ---------
|
16
|
+
# A+C B+D N = A+B+C+D
|
17
|
+
#
|
18
|
+
# P(f) = (A+B)/N
|
19
|
+
# P(f') = (C+D)/N
|
20
|
+
# P(c) = (A+C)/N
|
21
|
+
# P(c') = (B+D)/N
|
22
|
+
# P(f,c) = A/N
|
23
|
+
# P(f,c') = B/N
|
24
|
+
# P(f',c) = C/N
|
25
|
+
# P(f',c') = D/N
|
26
|
+
# P(f|c) = A/(A+C)
|
27
|
+
# P(f|c') = B/(B+D)
|
28
|
+
# P(f'|c) = C/(A+C)
|
29
|
+
# P(f'|c') = D/(B+D)
|
30
|
+
#
|
31
|
+
class BaseDiscrete < Base
|
32
|
+
# initialize from an existing data structure
|
33
|
+
def initialize(data=nil)
|
34
|
+
super(data)
|
35
|
+
end
|
36
|
+
|
37
|
+
private
|
38
|
+
|
39
|
+
# count of sample (i.e. 'A') that
|
40
|
+
# contains feature (f) and belongs to class (k)
|
41
|
+
def get_A(f, k)
|
42
|
+
@A ||= calc_A
|
43
|
+
a = @A[k][f]
|
44
|
+
|
45
|
+
# add 0.5 to avoid any ZERO in denominator or numerator
|
46
|
+
a+=0.5 if a.zero?
|
47
|
+
|
48
|
+
a
|
49
|
+
end # get_A
|
50
|
+
|
51
|
+
|
52
|
+
# pre-compute 'A'
|
53
|
+
def calc_A
|
54
|
+
results = {}
|
55
|
+
|
56
|
+
each_class do |k1|
|
57
|
+
results[k1] = {}
|
58
|
+
|
59
|
+
each_feature do |f|
|
60
|
+
count = 0.0
|
61
|
+
|
62
|
+
each_sample do |k2, s|
|
63
|
+
if k2 == k1
|
64
|
+
count += 1 if s.has_key? f
|
65
|
+
end
|
66
|
+
end
|
67
|
+
|
68
|
+
results[k1][f] = count
|
69
|
+
end
|
70
|
+
end
|
71
|
+
|
72
|
+
results
|
73
|
+
end # calc_A
|
74
|
+
|
75
|
+
|
76
|
+
# count of sample (i.e. 'B') that
|
77
|
+
# contains feature (f) but does not belong to class (k)
|
78
|
+
def get_B(f, k)
|
79
|
+
@B ||= calc_B
|
80
|
+
b = @B[k][f]
|
81
|
+
|
82
|
+
# add 0.5 to avoid any ZERO in denominator or numerator
|
83
|
+
b+=0.5 if b.zero?
|
84
|
+
|
85
|
+
b
|
86
|
+
end # get_B
|
87
|
+
|
88
|
+
|
89
|
+
# pre-compute 'B'
|
90
|
+
def calc_B
|
91
|
+
results = {}
|
92
|
+
|
93
|
+
each_class do |k1|
|
94
|
+
results[k1] = {}
|
95
|
+
|
96
|
+
each_feature do |f|
|
97
|
+
count = 0.0
|
98
|
+
|
99
|
+
each_sample do |k2, s|
|
100
|
+
if k2 != k1
|
101
|
+
count += 1 if s.has_key? f
|
102
|
+
end
|
103
|
+
end
|
104
|
+
|
105
|
+
results[k1][f] = count
|
106
|
+
end
|
107
|
+
end
|
108
|
+
|
109
|
+
results
|
110
|
+
end # calc_B
|
111
|
+
|
112
|
+
|
113
|
+
# count of sample (i.e. 'C') that
|
114
|
+
# does not contain feature (f) but belongs to class (k)
|
115
|
+
def get_C(f, k)
|
116
|
+
@C ||= calc_C
|
117
|
+
c = @C[k][f]
|
118
|
+
|
119
|
+
# add 0.5 to avoid any ZERO in denominator or numerator
|
120
|
+
c+=0.5 if c.zero?
|
121
|
+
|
122
|
+
c
|
123
|
+
end # get_C
|
124
|
+
|
125
|
+
|
126
|
+
# pre-compute 'C'
|
127
|
+
def calc_C
|
128
|
+
results = {}
|
129
|
+
|
130
|
+
each_class do |k1|
|
131
|
+
results[k1] = {}
|
132
|
+
|
133
|
+
each_feature do |f|
|
134
|
+
count = 0.0
|
135
|
+
|
136
|
+
each_sample do |k2, s|
|
137
|
+
if k2 == k1
|
138
|
+
count += 1 if not s.has_key? f
|
139
|
+
end
|
140
|
+
end
|
141
|
+
|
142
|
+
results[k1][f] = count
|
143
|
+
end
|
144
|
+
end
|
145
|
+
|
146
|
+
results
|
147
|
+
end # calc_C
|
148
|
+
|
149
|
+
|
150
|
+
# count of sample (i.e. 'D') that
|
151
|
+
# does not contain feature (f) and does not belong to class (c)
|
152
|
+
def get_D(f, k)
|
153
|
+
@D ||= calc_D
|
154
|
+
d = @D[k][f]
|
155
|
+
|
156
|
+
# add 0.5 to avoid any ZERO in denominator or numerator
|
157
|
+
d+=0.5 if d.zero?
|
158
|
+
|
159
|
+
d
|
160
|
+
end # get_D
|
161
|
+
|
162
|
+
|
163
|
+
# pre-compute 'D'
|
164
|
+
def calc_D
|
165
|
+
results = {}
|
166
|
+
|
167
|
+
each_class do |k1|
|
168
|
+
results[k1] = {}
|
169
|
+
|
170
|
+
each_feature do |f|
|
171
|
+
count = 0.0
|
172
|
+
|
173
|
+
each_sample do |k2, s|
|
174
|
+
if k2 != k1
|
175
|
+
count += 1 if not s.has_key? f
|
176
|
+
end
|
177
|
+
end
|
178
|
+
|
179
|
+
results[k1][f] = count
|
180
|
+
end
|
181
|
+
end
|
182
|
+
|
183
|
+
results
|
184
|
+
end # calc_D
|
185
|
+
|
186
|
+
|
187
|
+
end # class
|
188
|
+
|
189
|
+
|
190
|
+
end # module
|
@@ -0,0 +1,47 @@
|
|
1
|
+
#
|
2
|
+
# FSelector: a Ruby gem for feature selection and ranking
|
3
|
+
#
|
4
|
+
module FSelector
|
5
|
+
#
|
6
|
+
# Correlation-based Feature Selection (CFS) algorithm for continuous feature (CFS_c)
|
7
|
+
#
|
8
|
+
# ref: [Feature Selection for Discrete and Numeric Class Machine Learning](http://www.cs.waikato.ac.nz/ml/publications/1999/99MH-Feature-Select.pdf)
|
9
|
+
#
|
10
|
+
class CFS_c < BaseCFS
|
11
|
+
|
12
|
+
private
|
13
|
+
|
14
|
+
# calc the feature-class correlation of two vectors
|
15
|
+
def do_rcf(cv, fv)
|
16
|
+
# weighted pearson's correlation as cv (class label) contains discrete data
|
17
|
+
r = 0.0
|
18
|
+
|
19
|
+
cv.uniq.each do |k|
|
20
|
+
v = []
|
21
|
+
p = cv.count(k)/cv.size.to_f
|
22
|
+
|
23
|
+
cv.each do |c|
|
24
|
+
if c == k
|
25
|
+
v << 1
|
26
|
+
else
|
27
|
+
v << 0
|
28
|
+
end
|
29
|
+
end
|
30
|
+
|
31
|
+
r += p*v.pearson_r(fv)
|
32
|
+
end
|
33
|
+
|
34
|
+
r
|
35
|
+
end # do_rcf
|
36
|
+
|
37
|
+
|
38
|
+
# calc the feature-feature correlation of two vectors
|
39
|
+
def do_rff(fv, sv)
|
40
|
+
fv.pearson_r(sv) # use pearson's correlation coefficient
|
41
|
+
end # do_rff
|
42
|
+
|
43
|
+
|
44
|
+
end # class
|
45
|
+
|
46
|
+
|
47
|
+
end # module
|
@@ -9,90 +9,12 @@ module FSelector
|
|
9
9
|
#
|
10
10
|
# ref: [Estimating Attributes: Analysis and Extensions of RELIEF](http://www.springerlink.com/content/fp23jh2h0426ww45/)
|
11
11
|
#
|
12
|
-
class ReliefF_c <
|
13
|
-
#
|
14
|
-
# new()
|
15
|
-
#
|
16
|
-
# @param [Integer] m number of samples to be used
|
17
|
-
# for estimating feature contribution. max can be
|
18
|
-
# the number of training samples
|
19
|
-
# @param [Integer] k number of k-nearest neighbor
|
20
|
-
# @param [Hash] data existing data structure
|
21
|
-
#
|
22
|
-
def initialize(m=nil, k=10, data=nil)
|
23
|
-
super(data)
|
24
|
-
@m = m # use all samples
|
25
|
-
@k = (k || 10) # default 10
|
26
|
-
end
|
27
|
-
|
28
|
-
private
|
29
|
-
|
30
|
-
# calculate contribution of each feature (f) across all classes
|
31
|
-
def calc_contribution(f)
|
32
|
-
score = 0.0
|
33
|
-
|
34
|
-
# use all samples if @m not provided
|
35
|
-
@m = get_sample_size if not @m
|
36
|
-
|
37
|
-
@m.times do
|
38
|
-
# pick a sample at random
|
39
|
-
rs, rk = pick_a_sample_at_random
|
40
|
-
|
41
|
-
# find k nearest neighbor for each class
|
42
|
-
nbrs = find_k_nearest_nb(rs, rk)
|
43
|
-
|
44
|
-
# calc contribution from neighbors
|
45
|
-
score += calc_score(f, rs, rk, nbrs)
|
46
|
-
end
|
47
|
-
|
48
|
-
s = score / @m
|
49
|
-
|
50
|
-
set_feature_score(f, :BEST, s)
|
51
|
-
end # calc_contribution
|
52
|
-
|
53
|
-
|
54
|
-
# pick a sample at random
|
55
|
-
def pick_a_sample_at_random
|
56
|
-
rk = get_classes[rand(get_classes.size)]
|
57
|
-
rks = get_data[rk]
|
58
|
-
|
59
|
-
[ rks[rand(rks.size)], rk ]
|
60
|
-
end # pick_a_sample_at_random
|
61
|
-
|
62
|
-
# # find k nearest neighbors of sample (rs) for each class
|
63
|
-
def find_k_nearest_nb(rs, rk)
|
64
|
-
nbrs = {}
|
65
|
-
|
66
|
-
each_class do |k|
|
67
|
-
res = []
|
68
|
-
|
69
|
-
get_data[k].each do |s|
|
70
|
-
next if s == rs # exclude self
|
71
|
-
|
72
|
-
d = diff_sample(rs, s, rk, k)
|
73
|
-
res << [d, s]
|
74
|
-
end
|
75
|
-
|
76
|
-
nbrs[k] = (res.sort { |x, y| x[0] <=> y[0] }[0...@k]).collect { |z| z[1] }
|
77
|
-
end
|
78
|
-
|
79
|
-
nbrs
|
80
|
-
end # find_k_nearest_nb
|
81
|
-
|
82
|
-
|
83
|
-
# difference between two samples
|
84
|
-
def diff_sample(s1, s2, k1, k2)
|
85
|
-
d = 0.0
|
86
|
-
|
87
|
-
each_feature do |f|
|
88
|
-
d += diff_feature(f, s1, s2, k1, k2)**2
|
89
|
-
end
|
90
|
-
|
91
|
-
d
|
92
|
-
end # diff_sample
|
12
|
+
class ReliefF_c < BaseReliefF
|
93
13
|
|
14
|
+
private
|
94
15
|
|
95
16
|
# difference beween the feature (f) of two samples
|
17
|
+
# specialized version for continuous feature
|
96
18
|
def diff_feature(f, s1, s2, k1, k2)
|
97
19
|
d = 0.0
|
98
20
|
|
@@ -117,36 +39,6 @@ module FSelector
|
|
117
39
|
end # diff_feature
|
118
40
|
|
119
41
|
|
120
|
-
# calc probability of missing value (mv)
|
121
|
-
def calc_p(f, mv, k)
|
122
|
-
# cache
|
123
|
-
if not @f2mvp
|
124
|
-
@f2mvp = {}
|
125
|
-
|
126
|
-
each_feature do |f|
|
127
|
-
@f2mvp[f] = {}
|
128
|
-
|
129
|
-
each_class do |k|
|
130
|
-
@f2mvp[f][k] = {}
|
131
|
-
|
132
|
-
fvs = get_feature_values(f).uniq
|
133
|
-
fvs.each do |v|
|
134
|
-
n = 0.0
|
135
|
-
|
136
|
-
get_data[k].each do |s|
|
137
|
-
n += 1 if s.has_key?(f) and s[f] == v
|
138
|
-
end
|
139
|
-
|
140
|
-
@f2mvp[f][k][v] = n/get_data[k].size
|
141
|
-
end
|
142
|
-
end
|
143
|
-
end
|
144
|
-
end
|
145
|
-
|
146
|
-
@f2mvp[f][k][mv]
|
147
|
-
end
|
148
|
-
|
149
|
-
|
150
42
|
# get normalization unit for each feature
|
151
43
|
def get_normalization_unit(fi)
|
152
44
|
return @f2nu[fi] if @f2nu
|
@@ -162,28 +54,7 @@ module FSelector
|
|
162
54
|
end # get_normalization_unit
|
163
55
|
|
164
56
|
|
165
|
-
# calc feature (f) contribution from neighbors
|
166
|
-
def calc_score(f, rs, rk, nbrs)
|
167
|
-
score = 0.0
|
168
|
-
|
169
|
-
nbrs.each do |k, nbs|
|
170
|
-
if k == rk # near hit
|
171
|
-
nbs.each do |s|
|
172
|
-
score -= (diff_feature(f, rs, s, rk, k)**2/nbs.size)
|
173
|
-
end
|
174
|
-
else # near_miss
|
175
|
-
nbs.each do |s|
|
176
|
-
score += (get_data[k].size/get_sample_size.to_f *
|
177
|
-
diff_feature(f, rs, s, rk, k)**2/nbs.size)
|
178
|
-
end
|
179
|
-
end
|
180
|
-
end
|
181
|
-
|
182
|
-
score
|
183
|
-
end
|
184
|
-
|
185
|
-
|
186
57
|
end # class
|
187
58
|
|
188
59
|
|
189
|
-
end # module
|
60
|
+
end # module
|