1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
|
package de.lmu.ifi.dbs.elki.math.linearalgebra.pca;
/*
This file is part of ELKI:
Environment for Developing KDD-Applications Supported by Index-Structures
Copyright (C) 2012
Ludwig-Maximilians-Universität München
Lehr- und Forschungseinheit für Datenbanksysteme
ELKI Development Team
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
import java.util.ArrayList;
import java.util.List;
import de.lmu.ifi.dbs.elki.math.linearalgebra.EigenPair;
import de.lmu.ifi.dbs.elki.math.linearalgebra.SortedEigenPairs;
import de.lmu.ifi.dbs.elki.utilities.documentation.Description;
import de.lmu.ifi.dbs.elki.utilities.documentation.Title;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.AbstractParameterizer;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.OptionID;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.constraints.GreaterConstraint;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.constraints.GreaterEqualConstraint;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.constraints.LessConstraint;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.parameterization.Parameterization;
import de.lmu.ifi.dbs.elki.utilities.optionhandling.parameters.DoubleParameter;
/**
* The ProgressiveEigenPairFilter sorts the eigenpairs in descending order of
* their eigenvalues and marks the first eigenpairs, whose sum of eigenvalues is
* higher than the given percentage of the sum of all eigenvalues as strong
* eigenpairs. In contrast to the PercentageEigenPairFilter, it will use a
* percentage which changes linearly with the subspace dimensionality. This
* makes the parameter more consistent for different dimensionalities and often
* gives better results when clusters of different dimensionality exist, since
* different percentage alpha levels might be appropriate for different
* dimensionalities.
*
* Example calculations of alpha levels:
*
* In a 3D space, a progressive alpha value of 0.5 equals:
*
* - 1D subspace: 50 % + 1/3 of remainder = 0.667
*
* - 2D subspace: 50 % + 2/3 of remainder = 0.833
*
* In a 4D space, a progressive alpha value of 0.5 equals:
*
* - 1D subspace: 50% + 1/4 of remainder = 0.625
*
* - 2D subspace: 50% + 2/4 of remainder = 0.750
*
* - 3D subspace: 50% + 3/4 of remainder = 0.875
*
* Reasoning why this improves over PercentageEigenPairFilter:
*
* In a 100 dimensional space, a single Eigenvector representing over 85% of the
* total variance is highly significant, whereas the strongest 85 Eigenvectors
* together will by definition always represent at least 85% of the variance.
* PercentageEigenPairFilter can thus not be used with these parameters and
* detect both dimensionalities correctly.
*
* The second parameter introduced here, walpha, serves a different function: It
* prevents the eigenpair filter to use a statistically weak Eigenvalue just to
* reach the intended level, e.g. 84% + 1% >= 85% when 1% is statistically very
* weak.
*
* @author Erich Schubert
*
*/
@Title("Progressive Eigenpair Filter")
@Description("Sorts the eigenpairs in decending order of their eigenvalues and returns the first eigenpairs, whose sum of eigenvalues explains more than the a certain percentage of the unexpected variance, where the percentage increases with subspace dimensionality.")
public class ProgressiveEigenPairFilter implements EigenPairFilter {
/**
* Parameter progressive alpha.
*/
public static final OptionID EIGENPAIR_FILTER_PALPHA = new OptionID("pca.filter.progressivealpha", "The share (0.0 to 1.0) of variance that needs to be explained by the 'strong' eigenvectors." + "The filter class will choose the number of strong eigenvectors by this share.");
/**
* The default value for alpha.
*/
public static final double DEFAULT_PALPHA = 0.5;
/**
* The default value for alpha.
*/
public static final double DEFAULT_WALPHA = 0.95;
/**
* The threshold for strong eigenvectors: the strong eigenvectors explain a
* portion of at least alpha of the total variance.
*/
private double palpha;
/**
* The noise tolerance level for weak eigenvectors
*/
private double walpha;
/**
* Constructor.
*
* @param palpha palpha
* @param walpha walpha
*/
public ProgressiveEigenPairFilter(double palpha, double walpha) {
super();
this.palpha = palpha;
this.walpha = walpha;
}
/**
* Filter eigenpairs.
*/
@Override
public FilteredEigenPairs filter(SortedEigenPairs eigenPairs) {
// init strong and weak eigenpairs
List<EigenPair> strongEigenPairs = new ArrayList<EigenPair>();
List<EigenPair> weakEigenPairs = new ArrayList<EigenPair>();
// determine sum of eigenvalues
double totalSum = 0;
for (int i = 0; i < eigenPairs.size(); i++) {
EigenPair eigenPair = eigenPairs.getEigenPair(i);
totalSum += eigenPair.getEigenvalue();
}
double expectedVariance = totalSum / eigenPairs.size() * walpha;
// determine strong and weak eigenpairs
double currSum = 0;
boolean found = false;
int i;
for (i = 0; i < eigenPairs.size(); i++) {
EigenPair eigenPair = eigenPairs.getEigenPair(i);
// weak Eigenvector?
if (eigenPair.getEigenvalue() < expectedVariance) {
break;
}
currSum += eigenPair.getEigenvalue();
// calculate progressive alpha level
double alpha = 1.0 - (1.0 - palpha) * (1.0 - (i + 1) / eigenPairs.size());
if (currSum / totalSum >= alpha || i == eigenPairs.size() - 1) {
found = true;
strongEigenPairs.add(eigenPair);
break;
}
}
// if we didn't hit our alpha level, we consider all vectors to be weak!
if (!found) {
assert (weakEigenPairs.size() == 0);
weakEigenPairs = strongEigenPairs;
strongEigenPairs = new ArrayList<EigenPair>();
}
for (; i < eigenPairs.size(); i++) {
EigenPair eigenPair = eigenPairs.getEigenPair(i);
weakEigenPairs.add(eigenPair);
}
// the code using this method doesn't expect an empty strong set,
// if we didn't find any strong ones, we make all vectors strong
if (strongEigenPairs.size() == 0) {
return new FilteredEigenPairs(new ArrayList<EigenPair>(), weakEigenPairs);
}
return new FilteredEigenPairs(weakEigenPairs, strongEigenPairs);
}
/**
* Parameterization class.
*
* @author Erich Schubert
*
* @apiviz.exclude
*/
public static class Parameterizer extends AbstractParameterizer {
/**
* The threshold for strong eigenvectors: the strong eigenvectors explain a
* portion of at least alpha of the total variance.
*/
private double palpha;
/**
* The noise tolerance level for weak eigenvectors
*/
private double walpha;
@Override
protected void makeOptions(Parameterization config) {
super.makeOptions(config);
DoubleParameter palphaP = new DoubleParameter(EIGENPAIR_FILTER_PALPHA, DEFAULT_PALPHA);
palphaP.addConstraint(new GreaterConstraint(0.0));
palphaP.addConstraint(new LessConstraint(1.0));
if (config.grab(palphaP)) {
palpha = palphaP.getValue();
}
DoubleParameter walphaP = new DoubleParameter(WeakEigenPairFilter.EIGENPAIR_FILTER_WALPHA, DEFAULT_WALPHA);
walphaP.addConstraint(new GreaterEqualConstraint(0.0));
if (config.grab(walphaP)) {
walpha = walphaP.getValue();
}
}
@Override
protected ProgressiveEigenPairFilter makeInstance() {
return new ProgressiveEigenPairFilter(palpha, walpha);
}
}
}
|
