[3839] | 1 | /*************************************************************************
|
---|
| 2 | Copyright (c) 2008, Sergey Bochkanov (ALGLIB project).
|
---|
| 3 |
|
---|
| 4 | >>> SOURCE LICENSE >>>
|
---|
| 5 | This program is free software; you can redistribute it and/or modify
|
---|
| 6 | it under the terms of the GNU General Public License as published by
|
---|
| 7 | the Free Software Foundation (www.fsf.org); either version 2 of the
|
---|
| 8 | License, or (at your option) any later version.
|
---|
| 9 |
|
---|
| 10 | This program is distributed in the hope that it will be useful,
|
---|
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
| 13 | GNU General Public License for more details.
|
---|
| 14 |
|
---|
| 15 | A copy of the GNU General Public License is available at
|
---|
| 16 | http://www.fsf.org/licensing/licenses
|
---|
| 17 |
|
---|
| 18 | >>> END OF LICENSE >>>
|
---|
| 19 | *************************************************************************/
|
---|
| 20 |
|
---|
| 21 | using System;
|
---|
| 22 |
|
---|
| 23 | namespace alglib
|
---|
| 24 | {
|
---|
| 25 | public class lda
|
---|
| 26 | {
|
---|
| 27 | /*************************************************************************
|
---|
| 28 | Multiclass Fisher LDA
|
---|
| 29 |
|
---|
| 30 | Subroutine finds coefficients of linear combination which optimally separates
|
---|
| 31 | training set on classes.
|
---|
| 32 |
|
---|
| 33 | INPUT PARAMETERS:
|
---|
| 34 | XY - training set, array[0..NPoints-1,0..NVars].
|
---|
| 35 | First NVars columns store values of independent
|
---|
| 36 | variables, next column stores number of class (from 0
|
---|
| 37 | to NClasses-1) which dataset element belongs to. Fractional
|
---|
| 38 | values are rounded to nearest integer.
|
---|
| 39 | NPoints - training set size, NPoints>=0
|
---|
| 40 | NVars - number of independent variables, NVars>=1
|
---|
| 41 | NClasses - number of classes, NClasses>=2
|
---|
| 42 |
|
---|
| 43 |
|
---|
| 44 | OUTPUT PARAMETERS:
|
---|
| 45 | Info - return code:
|
---|
| 46 | * -4, if internal EVD subroutine hasn't converged
|
---|
| 47 | * -2, if there is a point with class number
|
---|
| 48 | outside of [0..NClasses-1].
|
---|
| 49 | * -1, if incorrect parameters was passed (NPoints<0,
|
---|
| 50 | NVars<1, NClasses<2)
|
---|
| 51 | * 1, if task has been solved
|
---|
| 52 | * 2, if there was a multicollinearity in training set,
|
---|
| 53 | but task has been solved.
|
---|
| 54 | W - linear combination coefficients, array[0..NVars-1]
|
---|
| 55 |
|
---|
| 56 | -- ALGLIB --
|
---|
| 57 | Copyright 31.05.2008 by Bochkanov Sergey
|
---|
| 58 | *************************************************************************/
|
---|
| 59 | public static void fisherlda(ref double[,] xy,
|
---|
| 60 | int npoints,
|
---|
| 61 | int nvars,
|
---|
| 62 | int nclasses,
|
---|
| 63 | ref int info,
|
---|
| 64 | ref double[] w)
|
---|
| 65 | {
|
---|
| 66 | double[,] w2 = new double[0,0];
|
---|
| 67 | int i_ = 0;
|
---|
| 68 |
|
---|
| 69 | fisherldan(ref xy, npoints, nvars, nclasses, ref info, ref w2);
|
---|
| 70 | if( info>0 )
|
---|
| 71 | {
|
---|
| 72 | w = new double[nvars-1+1];
|
---|
| 73 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 74 | {
|
---|
| 75 | w[i_] = w2[i_,0];
|
---|
| 76 | }
|
---|
| 77 | }
|
---|
| 78 | }
|
---|
| 79 |
|
---|
| 80 |
|
---|
| 81 | /*************************************************************************
|
---|
| 82 | N-dimensional multiclass Fisher LDA
|
---|
| 83 |
|
---|
| 84 | Subroutine finds coefficients of linear combinations which optimally separates
|
---|
| 85 | training set on classes. It returns N-dimensional basis whose vector are sorted
|
---|
| 86 | by quality of training set separation (in descending order).
|
---|
| 87 |
|
---|
| 88 | INPUT PARAMETERS:
|
---|
| 89 | XY - training set, array[0..NPoints-1,0..NVars].
|
---|
| 90 | First NVars columns store values of independent
|
---|
| 91 | variables, next column stores number of class (from 0
|
---|
| 92 | to NClasses-1) which dataset element belongs to. Fractional
|
---|
| 93 | values are rounded to nearest integer.
|
---|
| 94 | NPoints - training set size, NPoints>=0
|
---|
| 95 | NVars - number of independent variables, NVars>=1
|
---|
| 96 | NClasses - number of classes, NClasses>=2
|
---|
| 97 |
|
---|
| 98 |
|
---|
| 99 | OUTPUT PARAMETERS:
|
---|
| 100 | Info - return code:
|
---|
| 101 | * -4, if internal EVD subroutine hasn't converged
|
---|
| 102 | * -2, if there is a point with class number
|
---|
| 103 | outside of [0..NClasses-1].
|
---|
| 104 | * -1, if incorrect parameters was passed (NPoints<0,
|
---|
| 105 | NVars<1, NClasses<2)
|
---|
| 106 | * 1, if task has been solved
|
---|
| 107 | * 2, if there was a multicollinearity in training set,
|
---|
| 108 | but task has been solved.
|
---|
| 109 | W - basis, array[0..NVars-1,0..NVars-1]
|
---|
| 110 | columns of matrix stores basis vectors, sorted by
|
---|
| 111 | quality of training set separation (in descending order)
|
---|
| 112 |
|
---|
| 113 | -- ALGLIB --
|
---|
| 114 | Copyright 31.05.2008 by Bochkanov Sergey
|
---|
| 115 | *************************************************************************/
|
---|
| 116 | public static void fisherldan(ref double[,] xy,
|
---|
| 117 | int npoints,
|
---|
| 118 | int nvars,
|
---|
| 119 | int nclasses,
|
---|
| 120 | ref int info,
|
---|
| 121 | ref double[,] w)
|
---|
| 122 | {
|
---|
| 123 | int i = 0;
|
---|
| 124 | int j = 0;
|
---|
| 125 | int k = 0;
|
---|
| 126 | int m = 0;
|
---|
| 127 | double v = 0;
|
---|
| 128 | int[] c = new int[0];
|
---|
| 129 | double[] mu = new double[0];
|
---|
| 130 | double[,] muc = new double[0,0];
|
---|
| 131 | int[] nc = new int[0];
|
---|
| 132 | double[,] sw = new double[0,0];
|
---|
| 133 | double[,] st = new double[0,0];
|
---|
| 134 | double[,] z = new double[0,0];
|
---|
| 135 | double[,] z2 = new double[0,0];
|
---|
| 136 | double[,] tm = new double[0,0];
|
---|
| 137 | double[,] sbroot = new double[0,0];
|
---|
| 138 | double[,] a = new double[0,0];
|
---|
| 139 | double[,] xyproj = new double[0,0];
|
---|
| 140 | double[,] wproj = new double[0,0];
|
---|
| 141 | double[] tf = new double[0];
|
---|
| 142 | double[] d = new double[0];
|
---|
| 143 | double[] d2 = new double[0];
|
---|
| 144 | double[] work = new double[0];
|
---|
| 145 | int i_ = 0;
|
---|
| 146 |
|
---|
| 147 |
|
---|
| 148 | //
|
---|
| 149 | // Test data
|
---|
| 150 | //
|
---|
| 151 | if( npoints<0 | nvars<1 | nclasses<2 )
|
---|
| 152 | {
|
---|
| 153 | info = -1;
|
---|
| 154 | return;
|
---|
| 155 | }
|
---|
| 156 | for(i=0; i<=npoints-1; i++)
|
---|
| 157 | {
|
---|
| 158 | if( (int)Math.Round(xy[i,nvars])<0 | (int)Math.Round(xy[i,nvars])>=nclasses )
|
---|
| 159 | {
|
---|
| 160 | info = -2;
|
---|
| 161 | return;
|
---|
| 162 | }
|
---|
| 163 | }
|
---|
| 164 | info = 1;
|
---|
| 165 |
|
---|
| 166 | //
|
---|
| 167 | // Special case: NPoints<=1
|
---|
| 168 | // Degenerate task.
|
---|
| 169 | //
|
---|
| 170 | if( npoints<=1 )
|
---|
| 171 | {
|
---|
| 172 | info = 2;
|
---|
| 173 | w = new double[nvars-1+1, nvars-1+1];
|
---|
| 174 | for(i=0; i<=nvars-1; i++)
|
---|
| 175 | {
|
---|
| 176 | for(j=0; j<=nvars-1; j++)
|
---|
| 177 | {
|
---|
| 178 | if( i==j )
|
---|
| 179 | {
|
---|
| 180 | w[i,j] = 1;
|
---|
| 181 | }
|
---|
| 182 | else
|
---|
| 183 | {
|
---|
| 184 | w[i,j] = 0;
|
---|
| 185 | }
|
---|
| 186 | }
|
---|
| 187 | }
|
---|
| 188 | return;
|
---|
| 189 | }
|
---|
| 190 |
|
---|
| 191 | //
|
---|
| 192 | // Prepare temporaries
|
---|
| 193 | //
|
---|
| 194 | tf = new double[nvars-1+1];
|
---|
| 195 | work = new double[Math.Max(nvars, npoints)+1];
|
---|
| 196 |
|
---|
| 197 | //
|
---|
| 198 | // Convert class labels from reals to integers (just for convenience)
|
---|
| 199 | //
|
---|
| 200 | c = new int[npoints-1+1];
|
---|
| 201 | for(i=0; i<=npoints-1; i++)
|
---|
| 202 | {
|
---|
| 203 | c[i] = (int)Math.Round(xy[i,nvars]);
|
---|
| 204 | }
|
---|
| 205 |
|
---|
| 206 | //
|
---|
| 207 | // Calculate class sizes and means
|
---|
| 208 | //
|
---|
| 209 | mu = new double[nvars-1+1];
|
---|
| 210 | muc = new double[nclasses-1+1, nvars-1+1];
|
---|
| 211 | nc = new int[nclasses-1+1];
|
---|
| 212 | for(j=0; j<=nvars-1; j++)
|
---|
| 213 | {
|
---|
| 214 | mu[j] = 0;
|
---|
| 215 | }
|
---|
| 216 | for(i=0; i<=nclasses-1; i++)
|
---|
| 217 | {
|
---|
| 218 | nc[i] = 0;
|
---|
| 219 | for(j=0; j<=nvars-1; j++)
|
---|
| 220 | {
|
---|
| 221 | muc[i,j] = 0;
|
---|
| 222 | }
|
---|
| 223 | }
|
---|
| 224 | for(i=0; i<=npoints-1; i++)
|
---|
| 225 | {
|
---|
| 226 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 227 | {
|
---|
| 228 | mu[i_] = mu[i_] + xy[i,i_];
|
---|
| 229 | }
|
---|
| 230 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 231 | {
|
---|
| 232 | muc[c[i],i_] = muc[c[i],i_] + xy[i,i_];
|
---|
| 233 | }
|
---|
| 234 | nc[c[i]] = nc[c[i]]+1;
|
---|
| 235 | }
|
---|
| 236 | for(i=0; i<=nclasses-1; i++)
|
---|
| 237 | {
|
---|
| 238 | v = (double)(1)/(double)(nc[i]);
|
---|
| 239 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 240 | {
|
---|
| 241 | muc[i,i_] = v*muc[i,i_];
|
---|
| 242 | }
|
---|
| 243 | }
|
---|
| 244 | v = (double)(1)/(double)(npoints);
|
---|
| 245 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 246 | {
|
---|
| 247 | mu[i_] = v*mu[i_];
|
---|
| 248 | }
|
---|
| 249 |
|
---|
| 250 | //
|
---|
| 251 | // Create ST matrix
|
---|
| 252 | //
|
---|
| 253 | st = new double[nvars-1+1, nvars-1+1];
|
---|
| 254 | for(i=0; i<=nvars-1; i++)
|
---|
| 255 | {
|
---|
| 256 | for(j=0; j<=nvars-1; j++)
|
---|
| 257 | {
|
---|
| 258 | st[i,j] = 0;
|
---|
| 259 | }
|
---|
| 260 | }
|
---|
| 261 | for(k=0; k<=npoints-1; k++)
|
---|
| 262 | {
|
---|
| 263 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 264 | {
|
---|
| 265 | tf[i_] = xy[k,i_];
|
---|
| 266 | }
|
---|
| 267 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 268 | {
|
---|
| 269 | tf[i_] = tf[i_] - mu[i_];
|
---|
| 270 | }
|
---|
| 271 | for(i=0; i<=nvars-1; i++)
|
---|
| 272 | {
|
---|
| 273 | v = tf[i];
|
---|
| 274 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 275 | {
|
---|
| 276 | st[i,i_] = st[i,i_] + v*tf[i_];
|
---|
| 277 | }
|
---|
| 278 | }
|
---|
| 279 | }
|
---|
| 280 |
|
---|
| 281 | //
|
---|
| 282 | // Create SW matrix
|
---|
| 283 | //
|
---|
| 284 | sw = new double[nvars-1+1, nvars-1+1];
|
---|
| 285 | for(i=0; i<=nvars-1; i++)
|
---|
| 286 | {
|
---|
| 287 | for(j=0; j<=nvars-1; j++)
|
---|
| 288 | {
|
---|
| 289 | sw[i,j] = 0;
|
---|
| 290 | }
|
---|
| 291 | }
|
---|
| 292 | for(k=0; k<=npoints-1; k++)
|
---|
| 293 | {
|
---|
| 294 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 295 | {
|
---|
| 296 | tf[i_] = xy[k,i_];
|
---|
| 297 | }
|
---|
| 298 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 299 | {
|
---|
| 300 | tf[i_] = tf[i_] - muc[c[k],i_];
|
---|
| 301 | }
|
---|
| 302 | for(i=0; i<=nvars-1; i++)
|
---|
| 303 | {
|
---|
| 304 | v = tf[i];
|
---|
| 305 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 306 | {
|
---|
| 307 | sw[i,i_] = sw[i,i_] + v*tf[i_];
|
---|
| 308 | }
|
---|
| 309 | }
|
---|
| 310 | }
|
---|
| 311 |
|
---|
| 312 | //
|
---|
| 313 | // Maximize ratio J=(w'*ST*w)/(w'*SW*w).
|
---|
| 314 | //
|
---|
| 315 | // First, make transition from w to v such that w'*ST*w becomes v'*v:
|
---|
| 316 | // v = root(ST)*w = R*w
|
---|
| 317 | // R = root(D)*Z'
|
---|
| 318 | // w = (root(ST)^-1)*v = RI*v
|
---|
| 319 | // RI = Z*inv(root(D))
|
---|
| 320 | // J = (v'*v)/(v'*(RI'*SW*RI)*v)
|
---|
| 321 | // ST = Z*D*Z'
|
---|
| 322 | //
|
---|
| 323 | // so we have
|
---|
| 324 | //
|
---|
| 325 | // J = (v'*v) / (v'*(inv(root(D))*Z'*SW*Z*inv(root(D)))*v) =
|
---|
| 326 | // = (v'*v) / (v'*A*v)
|
---|
| 327 | //
|
---|
| 328 | if( !evd.smatrixevd(st, nvars, 1, true, ref d, ref z) )
|
---|
| 329 | {
|
---|
| 330 | info = -4;
|
---|
| 331 | return;
|
---|
| 332 | }
|
---|
| 333 | w = new double[nvars-1+1, nvars-1+1];
|
---|
| 334 | if( (double)(d[nvars-1])<=(double)(0) | (double)(d[0])<=(double)(1000*AP.Math.MachineEpsilon*d[nvars-1]) )
|
---|
| 335 | {
|
---|
| 336 |
|
---|
| 337 | //
|
---|
| 338 | // Special case: D[NVars-1]<=0
|
---|
| 339 | // Degenerate task (all variables takes the same value).
|
---|
| 340 | //
|
---|
| 341 | if( (double)(d[nvars-1])<=(double)(0) )
|
---|
| 342 | {
|
---|
| 343 | info = 2;
|
---|
| 344 | for(i=0; i<=nvars-1; i++)
|
---|
| 345 | {
|
---|
| 346 | for(j=0; j<=nvars-1; j++)
|
---|
| 347 | {
|
---|
| 348 | if( i==j )
|
---|
| 349 | {
|
---|
| 350 | w[i,j] = 1;
|
---|
| 351 | }
|
---|
| 352 | else
|
---|
| 353 | {
|
---|
| 354 | w[i,j] = 0;
|
---|
| 355 | }
|
---|
| 356 | }
|
---|
| 357 | }
|
---|
| 358 | return;
|
---|
| 359 | }
|
---|
| 360 |
|
---|
| 361 | //
|
---|
| 362 | // Special case: degenerate ST matrix, multicollinearity found.
|
---|
| 363 | // Since we know ST eigenvalues/vectors we can translate task to
|
---|
| 364 | // non-degenerate form.
|
---|
| 365 | //
|
---|
| 366 | // Let WG is orthogonal basis of the non zero variance subspace
|
---|
| 367 | // of the ST and let WZ is orthogonal basis of the zero variance
|
---|
| 368 | // subspace.
|
---|
| 369 | //
|
---|
| 370 | // Projection on WG allows us to use LDA on reduced M-dimensional
|
---|
| 371 | // subspace, N-M vectors of WZ allows us to update reduced LDA
|
---|
| 372 | // factors to full N-dimensional subspace.
|
---|
| 373 | //
|
---|
| 374 | m = 0;
|
---|
| 375 | for(k=0; k<=nvars-1; k++)
|
---|
| 376 | {
|
---|
| 377 | if( (double)(d[k])<=(double)(1000*AP.Math.MachineEpsilon*d[nvars-1]) )
|
---|
| 378 | {
|
---|
| 379 | m = k+1;
|
---|
| 380 | }
|
---|
| 381 | }
|
---|
| 382 | System.Diagnostics.Debug.Assert(m!=0, "FisherLDAN: internal error #1");
|
---|
| 383 | xyproj = new double[npoints-1+1, nvars-m+1];
|
---|
| 384 | blas.matrixmatrixmultiply(ref xy, 0, npoints-1, 0, nvars-1, false, ref z, 0, nvars-1, m, nvars-1, false, 1.0, ref xyproj, 0, npoints-1, 0, nvars-m-1, 0.0, ref work);
|
---|
| 385 | for(i=0; i<=npoints-1; i++)
|
---|
| 386 | {
|
---|
| 387 | xyproj[i,nvars-m] = xy[i,nvars];
|
---|
| 388 | }
|
---|
| 389 | fisherldan(ref xyproj, npoints, nvars-m, nclasses, ref info, ref wproj);
|
---|
| 390 | if( info<0 )
|
---|
| 391 | {
|
---|
| 392 | return;
|
---|
| 393 | }
|
---|
| 394 | blas.matrixmatrixmultiply(ref z, 0, nvars-1, m, nvars-1, false, ref wproj, 0, nvars-m-1, 0, nvars-m-1, false, 1.0, ref w, 0, nvars-1, 0, nvars-m-1, 0.0, ref work);
|
---|
| 395 | for(k=nvars-m; k<=nvars-1; k++)
|
---|
| 396 | {
|
---|
| 397 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 398 | {
|
---|
| 399 | w[i_,k] = z[i_,k-(nvars-m)];
|
---|
| 400 | }
|
---|
| 401 | }
|
---|
| 402 | info = 2;
|
---|
| 403 | }
|
---|
| 404 | else
|
---|
| 405 | {
|
---|
| 406 |
|
---|
| 407 | //
|
---|
| 408 | // General case: no multicollinearity
|
---|
| 409 | //
|
---|
| 410 | tm = new double[nvars-1+1, nvars-1+1];
|
---|
| 411 | a = new double[nvars-1+1, nvars-1+1];
|
---|
| 412 | blas.matrixmatrixmultiply(ref sw, 0, nvars-1, 0, nvars-1, false, ref z, 0, nvars-1, 0, nvars-1, false, 1.0, ref tm, 0, nvars-1, 0, nvars-1, 0.0, ref work);
|
---|
| 413 | blas.matrixmatrixmultiply(ref z, 0, nvars-1, 0, nvars-1, true, ref tm, 0, nvars-1, 0, nvars-1, false, 1.0, ref a, 0, nvars-1, 0, nvars-1, 0.0, ref work);
|
---|
| 414 | for(i=0; i<=nvars-1; i++)
|
---|
| 415 | {
|
---|
| 416 | for(j=0; j<=nvars-1; j++)
|
---|
| 417 | {
|
---|
| 418 | a[i,j] = a[i,j]/Math.Sqrt(d[i]*d[j]);
|
---|
| 419 | }
|
---|
| 420 | }
|
---|
| 421 | if( !evd.smatrixevd(a, nvars, 1, true, ref d2, ref z2) )
|
---|
| 422 | {
|
---|
| 423 | info = -4;
|
---|
| 424 | return;
|
---|
| 425 | }
|
---|
| 426 | for(k=0; k<=nvars-1; k++)
|
---|
| 427 | {
|
---|
| 428 | for(i=0; i<=nvars-1; i++)
|
---|
| 429 | {
|
---|
| 430 | tf[i] = z2[i,k]/Math.Sqrt(d[i]);
|
---|
| 431 | }
|
---|
| 432 | for(i=0; i<=nvars-1; i++)
|
---|
| 433 | {
|
---|
| 434 | v = 0.0;
|
---|
| 435 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 436 | {
|
---|
| 437 | v += z[i,i_]*tf[i_];
|
---|
| 438 | }
|
---|
| 439 | w[i,k] = v;
|
---|
| 440 | }
|
---|
| 441 | }
|
---|
| 442 | }
|
---|
| 443 |
|
---|
| 444 | //
|
---|
| 445 | // Post-processing:
|
---|
| 446 | // * normalization
|
---|
| 447 | // * converting to non-negative form, if possible
|
---|
| 448 | //
|
---|
| 449 | for(k=0; k<=nvars-1; k++)
|
---|
| 450 | {
|
---|
| 451 | v = 0.0;
|
---|
| 452 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 453 | {
|
---|
| 454 | v += w[i_,k]*w[i_,k];
|
---|
| 455 | }
|
---|
| 456 | v = 1/Math.Sqrt(v);
|
---|
| 457 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 458 | {
|
---|
| 459 | w[i_,k] = v*w[i_,k];
|
---|
| 460 | }
|
---|
| 461 | v = 0;
|
---|
| 462 | for(i=0; i<=nvars-1; i++)
|
---|
| 463 | {
|
---|
| 464 | v = v+w[i,k];
|
---|
| 465 | }
|
---|
| 466 | if( (double)(v)<(double)(0) )
|
---|
| 467 | {
|
---|
| 468 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 469 | {
|
---|
| 470 | w[i_,k] = -1*w[i_,k];
|
---|
| 471 | }
|
---|
| 472 | }
|
---|
| 473 | }
|
---|
| 474 | }
|
---|
| 475 | }
|
---|
| 476 | }
|
---|