source: trunk/src/kNNI.cc @ 228

// $Id: kNNI.cc 228 2005-02-01 14:06:51Z peter $

#include "kNNI.h"

#include <algorithm>
#include <cmath>
#include <fstream>
#include <vector>

#include "stl_utility.h"

namespace theplu {
namespace cpptools {

  kNNI::kNNI(const gslapi::matrix& matrix,const gslapi::matrix& flag,
             const u_int neighbours)
    : NNI(matrix,flag,neighbours)
  {
    for (unsigned int i=0; i<weight_.rows(); i++)
      for (unsigned int j=0; j<weight_.columns(); j++)
        if (!weight_(i,j)) {
          mv_rows_.push_back(i);
          break;
        }
    //estimate();
  }



  // \hat{x_{ij}}=\frac{ \sum_{k=1,N} \frac{x_{kj}}{d_{ki}} }
  //                   { \sum_{k=1,N} \frac{1     }{d_{ki}} },
  // where N is defined in the paper cited in the NNI class definition
  // documentation.
  u_int kNNI::estimate(void)
  {
    using namespace std;
    for (unsigned int i=0; i<mv_rows_.size(); i++) {
      // Jari, avoid copying in next line
      vector<pair<u_int,double> > distance=calculate_distances(mv_rows_[i]);
      sort(distance.begin(),distance.end(),
                pair_value_compare<u_int,double>());
      for (unsigned int j=0; j<data_.columns(); j++)
        if (!weight_(mv_rows_[i],j)) {
          vector<u_int> knn=nearest_neighbours(j,distance);
          double new_value=0.0;
          double norm=0.0;
          for (vector<u_int>::const_iterator k=knn.begin(); k!=knn.end(); k++){
            // Jari, a small number needed here, use something standardized.
            // Avoid division with zero (perfect match vectors)
            double d=(distance[*k].second ? distance[*k].second : 1e-10);
            new_value+=data_(distance[*k].first,j)/d;
            norm+=1.0/d;
          }
          // No impute if no contributions from neighbours.
          if (norm)
            imputed_data_(mv_rows_[i],j)=new_value/norm;
          else
            not_imputed_.push_back(i);
        }
    }
    return not_imputed_.size();
  }


}} // of namespace cpptools and namespace theplu