vfem-engine.c

数据挖掘方面的源码

   int i, j;
   long n;
   ExamplePtr centroid, newCentroid, cMin, cMax, e;
   double denominator, numerator, delta, weight;
   double *denomValues;


   denomValues = MNewPtr(sizeof(double) * VALLength(is->centroids));

   DebugWarn(VALLength(is->centroids) == 0,
                  "newCentroid will be used uninitialized\n");
   newCentroid = 0;


   e = ExampleRead(boundData, es);
   n = 1;
   while(e != 0 && n <= is->n) {
      if(gMessageLevel > 4) {
         ExampleWrite(e, stdout);
      }

      /* HERE modify for negative Xs */
      /* do the W-Plusses */
      denominator = 0;

      for(i = 0 ; i < VALLength(is->centroids) ; i++) {
         centroid = VLIndex(is->centroids, i);
         cMax = VLIndex(is->cMax, i);
         cMin = VLIndex(is->cMin, i);

         denomValues[i] = exp( (-1.0 / (2.0 * gSigmaSquare)) *
           pow(AssignmentScaledDeltaMax(e, centroid, cMin, cMax, 
                            is->lastBound[i]), 2));
         denominator += denomValues[i];
      }

      for(i = 0 ; i < VALLength(is->centroids) ; i++) {
         centroid = VLIndex(is->centroids, i);
         newCentroid = VALIndex(newCentroids, i);
         cMax = VLIndex(is->cMax, i);
         cMin = VLIndex(is->cMin, i);

         numerator = exp( (-1.0 / (2.0 * gSigmaSquare)) *
            pow(AssignmentScaledDeltaMin(e, centroid, cMin, cMax,
                             is->lastBound[i]), 2));

         denominator -= denomValues[i];
         denominator += numerator;

         weight = (numerator /denominator);
         if(weight > 1.0) {
            weight = 1.0;
         }

         //printf("c%d num: %.4f denom: %.4f w+: %.4f\n", i, numerator,
         //        denominator, weight);
         //printf("   ob delta: %.4f as deltamin: %.4f\n",
         //        ExampleDistance(e, centroid),
         //         AssignmentScaledDeltaMin(e, centroid, cMin, cMax));

         is->wPlus[i] += weight;
         is->wPlusSquare[i] += pow(weight, 2);
         for(j = 0 ; j < ExampleSpecGetNumAttributes(es) ; j++) {
            is->wxPlus[i][j] += weight * 
                        ExampleGetContinuousAttributeValue(e, j);
 
            delta = ExampleGetContinuousAttributeValue(e, j) -
                    ExampleGetContinuousAttributeValue(newCentroid, j);
            if(delta > 0) {
               is->deltaPlus[i][j] += weight * delta;
            } else {
               is->deltaMinus[i][j] += weight * delta;
            }
            if(gMessageLevel > 4) {
               printf("w+ %.2f D %.2f D+ %.2f D- %.2f\n", 
                  weight, delta, is->deltaPlus[i][j], is->deltaMinus[i][j]);
            }
         }
         denominator += denomValues[i];
         denominator -= numerator;
      }

      /* HERE should denominator use min for the centroid in numerator? */
      /* do the W-Minuses */
      denominator = 0;
      for(i = 0 ; i < VALLength(is->centroids) ; i++) {
         centroid = VLIndex(is->centroids, i);
         cMax = VLIndex(is->cMax, i);
         cMin = VLIndex(is->cMin, i);

         denomValues[i] = exp( (-1.0 / (2.0 * gSigmaSquare)) *
            pow(AssignmentScaledDeltaMin(e, centroid, cMin, cMax,
                           is->lastBound[i]), 2));
         denominator += denomValues[i];
      } 
      for(i = 0 ; i < VALLength(is->centroids) ; i++) {
         centroid = VLIndex(is->centroids, i);
         cMax = VLIndex(is->cMax, i);
         cMin = VLIndex(is->cMin, i);

         numerator = exp( (-1.0 / (2.0 * gSigmaSquare)) *
            pow(AssignmentScaledDeltaMax(e, centroid, cMin, cMax,
                           is->lastBound[i]), 2));

         denominator -= denomValues[i];
         denominator += numerator;

         weight = (numerator /denominator);

         //printf("c%d num: %.4f denom: %.4f w-: %.4f\n", i, numerator,
         //          denominator, (numerator / denominator));
         //printf("   ob delta: %.4f as deltamax: %.4f\n",
         //          ExampleDistance(e, centroid),
         //         AssignmentScaledDeltaMax(e, centroid, cMin, cMax));

         is->wMinus[i] += weight;
         for(j = 0 ; j < ExampleSpecGetNumAttributes(es) ; j++) {
            is->wxMinus[i][j] += weight * 
                ExampleGetContinuousAttributeValue(e, j);

            delta = ExampleGetContinuousAttributeValue(e, j) -
                    ExampleGetContinuousAttributeValue(newCentroid, j);
            if(delta > 0) {
               is->deltaMinus[i][j] += weight * delta;
            } else {
               is->deltaPlus[i][j] += weight * delta;
            }
            if(gMessageLevel > 4) {
               printf("w- %.2f D %.2f D+ %.2f D- %.2f\n", 
                  weight, delta, is->deltaPlus[i][j], is->deltaMinus[i][j]);
            }
         }
         denominator += denomValues[i];
         denominator -= numerator;
      }

      ExampleFree(e);
      e = ExampleRead(boundData, es);
      n++;
   }

   if(e != 0) {
      ExampleFree(e);
   }

   for(i = 0 ; i < VALLength(is->centroids) ; i++) {
      for(j = 0 ; j < ExampleSpecGetNumAttributes(es) ; j++) {
         is->deltaPlus[i][j] = fabs(is->deltaPlus[i][j]);
         is->deltaMinus[i][j] = fabs(is->deltaMinus[i][j]);
      }
   }

   MFreePtr(denomValues);
}

float IterationStatsErrorBoundDimension(IterationStatsPtr is, int cluster,
					int dimension) {
   return max(
    fabs(
     ExampleGetContinuousAttributeValue(VALIndex(is->cMax, cluster),
                                                             dimension) -
     ExampleGetContinuousAttributeValue(VALIndex(is->centroids, cluster),
                                                dimension)),
    fabs(
     ExampleGetContinuousAttributeValue(VALIndex(is->cMin, cluster), 
                                                               dimension) -
     ExampleGetContinuousAttributeValue(VALIndex(is->centroids, cluster),
                                                        dimension)));

}

IterationStatsPtr IterationStatsNext(IterationStatsPtr is, float delta, float R, float assignErrorScale, ExampleSpecPtr es, int usePedroBound, FILE *boundData) {
   IterationStatsPtr newIs;
   int i, j;
   float thisBound;
   VoidAListPtr newCentroids = VALNew();

   /* initialize the new stat structure with the correct centroids */
   _MakeNewCentroids(is->centroids, newCentroids, (VoidAListPtr)0, 
        is->wx, is->w, is->wPlusSquare, is->wMinus, is->n, es,
         gNeededDelta, 0, 0, 0, (double **) 0, (double **) 0);
   newIs = IterationStatsInitial(newCentroids);
   newIs->isFirstIteration = 0;

   /* HERE MEM this leaks the memory of the newCentroids should example free */
   VALFree(newCentroids);

   if(usePedroBound) {
      _RecordPedroBoundInfo(boundData, is, newIs->centroids, es);

      if(gMessageLevel > 1) {
	IterationStatsWrite(is, es, stdout);
      }

      /* now make the min and max centroids */
      _MakeNewCentroids(is->centroids, newIs->cMax, newIs->cMaxAssignment,
           is->wx, is->w, is->wPlusSquare, is->wMinus, is->n, es, gNeededDelta,
                1, 1, !is->isFirstIteration, is->deltaPlus, is->deltaMinus);
      _MakeNewCentroids(is->centroids, newIs->cMin, newIs->cMinAssignment, 
           is->wx, is->w, is->wPlusSquare, is->wMinus, is->n, es, gNeededDelta,
                1, 0, !is->isFirstIteration, is->deltaPlus, is->deltaMinus);
   } else {
      /* now make the min and max centroids */
      _MakeNewCentroids(is->centroids, newIs->cMax, newIs->cMaxAssignment,
              is->wxPlus, is->wMinus,
         is->wPlusSquare, is->wMinus, is->n, es, gNeededDelta, 1, 1, 0, 
                   (double **)0, (double **)0);
      _MakeNewCentroids(is->centroids, newIs->cMin, newIs->cMinAssignment, 
              is->wxMinus, is->wPlus,
         is->wPlusSquare, is->wMinus, is->n, es, gNeededDelta, 1, 0, 0,
                    (double **)0, (double **)0);
   }

   newIs->foundBound = is->foundBound;

   /* now fill in the bounds information */
   for(i = 0 ; i < VALLength(is->centroids) ; i++) {
      if(is->w[i] == 0 || is->wMinus[i] == 0 || is->wPlus[i] == 0) {
         newIs->lastBound[i] = 1000000;
         newIs->lastAssignmentBound[i] = 1000000;
         newIs->foundBound = 0;
         //printf("*** n^ == n+ can't find a bound for this run\n");
         //exit(0);
      } else {
         newIs->lastBound[i] = max(
             ExampleDistance(VALIndex(newIs->centroids, i),
                                      VALIndex(newIs->cMax, i)),
             ExampleDistance(VALIndex(newIs->centroids, i),
                                      VALIndex(newIs->cMin, i)));

         newIs->lastAssignmentBound[i] = max (
             ExampleDistance(VALIndex(newIs->centroids, i),
                                      VALIndex(newIs->cMaxAssignment, i)),
             ExampleDistance(VALIndex(newIs->centroids, i),
                                      VALIndex(newIs->cMinAssignment, i)));

         for(j = 0 ; j < ExampleSpecGetNumAttributes(es) ; j++) {
            /* Find the largest dimension bound so we know
                 how many samples for next round if needed */
            /* HERE discrete? */
            thisBound = IterationStatsErrorBoundDimension(newIs, i, j);

            if(newIs->maxEkd < thisBound) {
               newIs->maxEkd = thisBound;
            }
         }
      }
   }

   if(gMessageLevel > 2) {
      printf("Made the next IS, centroids:\n");
      for(i = 0 ; i < VALLength(newIs->centroids) ; i++) {
         printf("lastBound %d: %.4f lastAssignmentBound: %.4f\n", i, 
                 newIs->lastBound[i], newIs->lastAssignmentBound[i]);
         printf("nCMin   %d: ", i);
         ExampleWrite(VALIndex(newIs->cMin, i), stdout);
         printf("nCMinAs %d: ", i);
         ExampleWrite(VALIndex(newIs->cMinAssignment, i), stdout);
         printf("nC      %d: ", i);
         ExampleWrite(VALIndex(newIs->centroids, i), stdout);
         printf("nCMaxAs %d: ", i);
         ExampleWrite(VALIndex(newIs->cMaxAssignment, i), stdout);
         printf("nCMax   %d: ", i);
         ExampleWrite(VALIndex(newIs->cMax, i), stdout);
         fflush(stdout);
      }
   }


   return newIs;
}

void IterationStatsWrite(IterationStatsPtr is, ExampleSpecPtr es, FILE *out) {
   int i, j;

   fprintf(out, "----------\n");
   for(i = 0 ; i < VALLength(is->centroids) ; i++) {
      fprintf(out, "center #%d lastBound: %.3lf lastAssignmentBound: %.3f\n", 
            i, is->lastBound[i], is->lastAssignmentBound[i]);
      fprintf(out, "   ");
      ExampleWrite(VALIndex(is->cMin, i), out);
      fprintf(out, "   ");
      ExampleWrite(VALIndex(is->centroids, i), out);
      fprintf(out, "   ");
      ExampleWrite(VALIndex(is->cMax, i), out);
      fprintf(out, "   w- %.2lf w %.2lf w+ %.2lf w+^2 %.2lf\n", is->wMinus[i], is->w[i], is->wPlus[i], is->wPlusSquare[i]);
      for(j = 0 ; j < ExampleSpecGetNumAttributes(es) ; j++) {
         fprintf(out, "      d%d: wx- %.2lf wx %.2lf wx+ %.2lf delta+ %.2f delta- %.2f\n", j, is->wxMinus[i][j], is->wx[i][j], is->wxPlus[i][j], is->deltaPlus[i][j], is->deltaMinus[i][j]);
      }
   }
}