1 | #ifndef _theplu_yat_classifier_ncc_ |
---|
2 | #define _theplu_yat_classifier_ncc_ |
---|
3 | |
---|
4 | // $Id$ |
---|
5 | |
---|
6 | /* |
---|
7 | Copyright (C) 2005 Peter Johansson, Markus Ringnér |
---|
8 | Copyright (C) 2006, 2007, 2008 Jari Häkkinen, Peter Johansson, Markus Ringnér |
---|
9 | Copyright (C) 2009, 2010 Peter Johansson |
---|
10 | |
---|
11 | This file is part of the yat library, http://dev.thep.lu.se/yat |
---|
12 | |
---|
13 | The yat library is free software; you can redistribute it and/or |
---|
14 | modify it under the terms of the GNU General Public License as |
---|
15 | published by the Free Software Foundation; either version 3 of the |
---|
16 | License, or (at your option) any later version. |
---|
17 | |
---|
18 | The yat library is distributed in the hope that it will be useful, |
---|
19 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
20 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
---|
21 | General Public License for more details. |
---|
22 | |
---|
23 | You should have received a copy of the GNU General Public License |
---|
24 | along with yat. If not, see <http://www.gnu.org/licenses/>. |
---|
25 | */ |
---|
26 | |
---|
27 | #include "MatrixLookup.h" |
---|
28 | #include "MatrixLookupWeighted.h" |
---|
29 | #include "SupervisedClassifier.h" |
---|
30 | #include "Target.h" |
---|
31 | |
---|
32 | #include "yat/statistics/Averager.h" |
---|
33 | #include "yat/statistics/AveragerWeighted.h" |
---|
34 | #include "yat/utility/concept_check.h" |
---|
35 | #include "yat/utility/Exception.h" |
---|
36 | #include "yat/utility/Matrix.h" |
---|
37 | #include "yat/utility/MatrixWeighted.h" |
---|
38 | #include "yat/utility/Vector.h" |
---|
39 | #include "yat/utility/stl_utility.h" |
---|
40 | #include "yat/utility/yat_assert.h" |
---|
41 | |
---|
42 | #include <boost/concept_check.hpp> |
---|
43 | |
---|
44 | #include <iterator> |
---|
45 | #include <map> |
---|
46 | #include <cmath> |
---|
47 | |
---|
48 | namespace theplu { |
---|
49 | namespace yat { |
---|
50 | namespace classifier { |
---|
51 | |
---|
52 | |
---|
53 | /** |
---|
54 | \brief Nearest Centroid Classifier |
---|
55 | |
---|
56 | A sample is predicted based on its distance to centroids for each |
---|
57 | class. The centroids are generated using training data. NCC |
---|
58 | supports using different measures, for example, Euclidean |
---|
59 | distance, to define distance between samples and centroids. |
---|
60 | |
---|
61 | The template argument Distance should be a class modelling |
---|
62 | the concept \ref concept_distance. |
---|
63 | */ |
---|
64 | template <typename Distance> |
---|
65 | class NCC : public SupervisedClassifier |
---|
66 | { |
---|
67 | |
---|
68 | public: |
---|
69 | /** |
---|
70 | \brief Constructor |
---|
71 | |
---|
72 | Distance is initialized using its default constructor. |
---|
73 | */ |
---|
74 | NCC(void); |
---|
75 | |
---|
76 | /** |
---|
77 | \brief Constructor using an initialized distance measure |
---|
78 | |
---|
79 | This constructor should be used if Distance has parameters and |
---|
80 | the user wants to specify the parameters by initializing |
---|
81 | Distance prior to constructing the NCC. |
---|
82 | */ |
---|
83 | NCC(const Distance&); |
---|
84 | |
---|
85 | |
---|
86 | /** |
---|
87 | Destructor |
---|
88 | */ |
---|
89 | virtual ~NCC(void); |
---|
90 | |
---|
91 | /** |
---|
92 | \brief Get the centroids for all classes. |
---|
93 | |
---|
94 | \return The centroids for each class as columns in a matrix. |
---|
95 | */ |
---|
96 | const utility::Matrix& centroids(void) const; |
---|
97 | |
---|
98 | NCC<Distance>* make_classifier(void) const; |
---|
99 | |
---|
100 | /** |
---|
101 | \brief Make predictions for unweighted test data. |
---|
102 | |
---|
103 | Predictions are calculated and returned in \a results. For |
---|
104 | each sample in \a data, \a results contains the distances to |
---|
105 | the centroids for each class. If a class has no training |
---|
106 | samples NaN's are returned for this class in \a |
---|
107 | results. Weighted distance calculations, in which NaN's have |
---|
108 | zero weights, are used if the centroids contain NaN's. |
---|
109 | |
---|
110 | \note NCC returns distances to centroids as the |
---|
111 | prediction. This means that the best class for a sample has the |
---|
112 | smallest value in \a results. This is in contrast to, for |
---|
113 | example, KNN for which the best class for a sample in \a |
---|
114 | results has the largest number (the largest number of nearest |
---|
115 | neighbors). |
---|
116 | */ |
---|
117 | void predict(const MatrixLookup& data, utility::Matrix& results) const; |
---|
118 | |
---|
119 | /** |
---|
120 | \brief Make predictions for weighted test data. |
---|
121 | |
---|
122 | Predictions are calculated and returned in \a results. For |
---|
123 | each sample in \a data, \a results contains the distances to |
---|
124 | the centroids for each class as in predict(const MatrixLookup& |
---|
125 | data, utility::Matrix& results). Weighted distance calculations |
---|
126 | are used, and zero weights are used for NaN's in centroids. If |
---|
127 | for a test sample and centroid pair, all variables have either |
---|
128 | zero weight for the test sample or NaN for the centroid, the |
---|
129 | centroid and the sample have no variables with values in |
---|
130 | common. In this case the prediction for the sample is set to |
---|
131 | NaN for the class in \a results. |
---|
132 | |
---|
133 | \note NCC returns distances to centroids as the |
---|
134 | prediction. This means that the best class for a sample has the |
---|
135 | smallest value in \a results. This is in contrast to, for |
---|
136 | example, KNN for which the best class for a sample in \a |
---|
137 | results has the largest number (the largest number of nearest |
---|
138 | neighbors). |
---|
139 | */ |
---|
140 | void predict(const MatrixLookupWeighted& data, utility::Matrix& results) const; |
---|
141 | |
---|
142 | /** |
---|
143 | \brief Train the NCC using unweighted training data with known |
---|
144 | targets. |
---|
145 | |
---|
146 | A centroid is calculated for each class. For each variable in |
---|
147 | \a data, a centroid for a class contains the average value of |
---|
148 | the variable across all training samples in the class. |
---|
149 | */ |
---|
150 | void train(const MatrixLookup& data, const Target& targets); |
---|
151 | |
---|
152 | |
---|
153 | /** |
---|
154 | \brief Train the NCC using weighted training data with known |
---|
155 | targets. |
---|
156 | |
---|
157 | A centroid is calculated for each class as in |
---|
158 | train(const MatrixLookup&, const Target&). |
---|
159 | The weights of the data are used when calculating the centroids |
---|
160 | and the centroids should be interpreted as unweighted |
---|
161 | (i.e. centroid values have unity weights). If a variable has |
---|
162 | zero weights for all samples in a class, the centroid is set to |
---|
163 | NaN for that variable. |
---|
164 | */ |
---|
165 | void train(const MatrixLookupWeighted& data, const Target& targets); |
---|
166 | |
---|
167 | |
---|
168 | private: |
---|
169 | |
---|
170 | void predict_unweighted(const MatrixLookup&, utility::Matrix&) const; |
---|
171 | void predict_weighted(const MatrixLookupWeighted&, utility::Matrix&) const; |
---|
172 | |
---|
173 | utility::Matrix centroids_; |
---|
174 | bool centroids_nan_; |
---|
175 | Distance distance_; |
---|
176 | }; |
---|
177 | |
---|
178 | // templates |
---|
179 | |
---|
180 | template <typename Distance> |
---|
181 | NCC<Distance>::NCC() |
---|
182 | : SupervisedClassifier(), centroids_nan_(false) |
---|
183 | { |
---|
184 | BOOST_CONCEPT_ASSERT((utility::DistanceConcept<Distance>)); |
---|
185 | } |
---|
186 | |
---|
187 | template <typename Distance> |
---|
188 | NCC<Distance>::NCC(const Distance& dist) |
---|
189 | : SupervisedClassifier(), centroids_nan_(false), distance_(dist) |
---|
190 | { |
---|
191 | BOOST_CONCEPT_ASSERT((utility::DistanceConcept<Distance>)); |
---|
192 | } |
---|
193 | |
---|
194 | |
---|
195 | template <typename Distance> |
---|
196 | NCC<Distance>::~NCC() |
---|
197 | { |
---|
198 | } |
---|
199 | |
---|
200 | |
---|
201 | template <typename Distance> |
---|
202 | const utility::Matrix& NCC<Distance>::centroids(void) const |
---|
203 | { |
---|
204 | return centroids_; |
---|
205 | } |
---|
206 | |
---|
207 | |
---|
208 | template <typename Distance> |
---|
209 | NCC<Distance>* |
---|
210 | NCC<Distance>::make_classifier() const |
---|
211 | { |
---|
212 | // All private members should be copied here to generate an |
---|
213 | // identical but untrained classifier |
---|
214 | return new NCC<Distance>(distance_); |
---|
215 | } |
---|
216 | |
---|
217 | template <typename Distance> |
---|
218 | void NCC<Distance>::train(const MatrixLookup& data, const Target& target) |
---|
219 | { |
---|
220 | centroids_.resize(data.rows(), target.nof_classes()); |
---|
221 | for(size_t i=0; i<data.rows(); i++) { |
---|
222 | std::vector<statistics::Averager> class_averager; |
---|
223 | class_averager.resize(target.nof_classes()); |
---|
224 | for(size_t j=0; j<data.columns(); j++) { |
---|
225 | class_averager[target(j)].add(data(i,j)); |
---|
226 | } |
---|
227 | for(size_t c=0;c<target.nof_classes();c++) { |
---|
228 | centroids_(i,c) = class_averager[c].mean(); |
---|
229 | } |
---|
230 | } |
---|
231 | } |
---|
232 | |
---|
233 | |
---|
234 | template <typename Distance> |
---|
235 | void NCC<Distance>::train(const MatrixLookupWeighted& data, const Target& target) |
---|
236 | { |
---|
237 | centroids_.resize(data.rows(), target.nof_classes()); |
---|
238 | for(size_t i=0; i<data.rows(); i++) { |
---|
239 | std::vector<statistics::AveragerWeighted> class_averager; |
---|
240 | class_averager.resize(target.nof_classes()); |
---|
241 | for(size_t j=0; j<data.columns(); j++) |
---|
242 | class_averager[target(j)].add(data.data(i,j),data.weight(i,j)); |
---|
243 | for(size_t c=0;c<target.nof_classes();c++) { |
---|
244 | if(class_averager[c].sum_w()==0) { |
---|
245 | centroids_nan_=true; |
---|
246 | } |
---|
247 | centroids_(i,c) = class_averager[c].mean(); |
---|
248 | } |
---|
249 | } |
---|
250 | } |
---|
251 | |
---|
252 | |
---|
253 | template <typename Distance> |
---|
254 | void NCC<Distance>::predict(const MatrixLookup& test, |
---|
255 | utility::Matrix& prediction) const |
---|
256 | { |
---|
257 | utility::yat_assert<utility::runtime_error> |
---|
258 | (centroids_.rows()==test.rows(), |
---|
259 | "NCC::predict test data with incorrect number of rows"); |
---|
260 | |
---|
261 | prediction.resize(centroids_.columns(), test.columns()); |
---|
262 | |
---|
263 | // If weighted training data has resulted in NaN in centroids: |
---|
264 | // weighted calculations |
---|
265 | if(centroids_nan_) { |
---|
266 | predict_weighted(MatrixLookupWeighted(test),prediction); |
---|
267 | } |
---|
268 | // If unweighted training data: unweighted calculations |
---|
269 | else { |
---|
270 | predict_unweighted(test,prediction); |
---|
271 | } |
---|
272 | } |
---|
273 | |
---|
274 | template <typename Distance> |
---|
275 | void NCC<Distance>::predict(const MatrixLookupWeighted& test, |
---|
276 | utility::Matrix& prediction) const |
---|
277 | { |
---|
278 | utility::yat_assert<utility::runtime_error> |
---|
279 | (centroids_.rows()==test.rows(), |
---|
280 | "NCC::predict test data with incorrect number of rows"); |
---|
281 | |
---|
282 | prediction.resize(centroids_.columns(), test.columns()); |
---|
283 | predict_weighted(test,prediction); |
---|
284 | } |
---|
285 | |
---|
286 | |
---|
287 | template <typename Distance> |
---|
288 | void NCC<Distance>::predict_unweighted(const MatrixLookup& test, |
---|
289 | utility::Matrix& prediction) const |
---|
290 | { |
---|
291 | for(size_t j=0; j<test.columns();j++) |
---|
292 | for(size_t k=0; k<centroids_.columns();k++) |
---|
293 | prediction(k,j) = distance_(test.begin_column(j), test.end_column(j), |
---|
294 | centroids_.begin_column(k)); |
---|
295 | } |
---|
296 | |
---|
297 | template <typename Distance> |
---|
298 | void NCC<Distance>::predict_weighted(const MatrixLookupWeighted& test, |
---|
299 | utility::Matrix& prediction) const |
---|
300 | { |
---|
301 | utility::MatrixWeighted weighted_centroids(centroids_); |
---|
302 | for(size_t j=0; j<test.columns();j++) |
---|
303 | for(size_t k=0; k<centroids_.columns();k++) |
---|
304 | prediction(k,j) = distance_(test.begin_column(j), test.end_column(j), |
---|
305 | weighted_centroids.begin_column(k)); |
---|
306 | } |
---|
307 | |
---|
308 | |
---|
309 | }}} // of namespace classifier, yat, and theplu |
---|
310 | |
---|
311 | #endif |
---|