void AverageLinkage::operator()(DistanceMatrix<float> & original_distance, std::vector<BinaryTreeNode> & cluster_tree, const float threshold /*=1*/) const
{
// input MUST have >= 2 elements!
if (original_distance.dimensionsize() < 2)
{
throw ClusterFunctor::InsufficientInput(__FILE__, __LINE__, OPENMS_PRETTY_FUNCTION, "Distance matrix to start from only contains one element");
}
std::vector<std::set<Size> > clusters(original_distance.dimensionsize());
for (Size i = 0; i < original_distance.dimensionsize(); ++i)
{
clusters[i].insert(i);
}
cluster_tree.clear();
cluster_tree.reserve(original_distance.dimensionsize() - 1);
// Initial minimum-distance pair
original_distance.updateMinElement();
std::pair<Size, Size> min = original_distance.getMinElementCoordinates();
Size overall_cluster_steps(original_distance.dimensionsize());
startProgress(0, original_distance.dimensionsize(), "clustering data");
while (original_distance(min.second, min.first) < threshold)
{
//grow the tree
cluster_tree.push_back(BinaryTreeNode(*(clusters[min.second].begin()), *(clusters[min.first].begin()), original_distance(min.first, min.second)));
if (cluster_tree.back().left_child > cluster_tree.back().right_child)
{
std::swap(cluster_tree.back().left_child, cluster_tree.back().right_child);
}
if (original_distance.dimensionsize() > 2)
{
//pick minimum-distance pair i,j and merge them
//calculate parameter for lance-williams formula
float alpha_i = (float)(clusters[min.first].size() / (float)(clusters[min.first].size() + clusters[min.second].size()));
float alpha_j = (float)(clusters[min.second].size() / (float)(clusters[min.first].size() + clusters[min.second].size()));
//~ std::cout << alpha_i << '\t' << alpha_j << std::endl;
//pushback elements of second to first (and then erase second)
clusters[min.second].insert(clusters[min.first].begin(), clusters[min.first].end());
// erase first one
clusters.erase(clusters.begin() + min.first);
//update original_distance matrix
//average linkage: new distance between clusters is the minimum distance between elements of each cluster
//lance-williams update for d((i,j),k): (m_i/m_i+m_j)* d(i,k) + (m_j/m_i+m_j)* d(j,k) ; m_x is the number of elements in cluster x
for (Size k = 0; k < min.second; ++k)
{
float dik = original_distance.getValue(min.first, k);
float djk = original_distance.getValue(min.second, k);
original_distance.setValueQuick(min.second, k, (alpha_i * dik + alpha_j * djk));
}
for (Size k = min.second + 1; k < original_distance.dimensionsize(); ++k)
{
float dik = original_distance.getValue(min.first, k);
float djk = original_distance.getValue(min.second, k);
original_distance.setValueQuick(k, min.second, (alpha_i * dik + alpha_j * djk));
}
//reduce
original_distance.reduce(min.first);
//update minimum-distance pair
original_distance.updateMinElement();
//get min-pair from triangular matrix
min = original_distance.getMinElementCoordinates();
}
else
{
break;
}
setProgress(overall_cluster_steps - original_distance.dimensionsize());
//repeat until only two cluster remains, last step skips matrix operations
}
//fill tree with dummy nodes
Size sad(*clusters.front().begin());
for (Size i = 1; (i < clusters.size()) && (cluster_tree.size() < cluster_tree.capacity()); ++i)
{
cluster_tree.push_back(BinaryTreeNode(sad, *clusters[i].begin(), -1.0));
}
endProgress();
}
void CompleteLinkage::operator()(DistanceMatrix<float> & original_distance, std::vector<BinaryTreeNode> & cluster_tree, const float threshold /*=1*/) const
{
// attention: clustering process is done by clustering the indices
// pointing to elements in inputvector and distances in inputmatrix
// input MUST have >= 2 elements!
if (original_distance.dimensionsize() < 2)
{
throw ClusterFunctor::InsufficientInput(__FILE__, __LINE__, OPENMS_PRETTY_FUNCTION, "Distance matrix to start from only contains one element");
}
std::vector<std::set<Size> > clusters(original_distance.dimensionsize());
for (Size i = 0; i < original_distance.dimensionsize(); ++i)
{
clusters[i].insert(i);
}
cluster_tree.clear();
cluster_tree.reserve(original_distance.dimensionsize() - 1);
// Initial minimum-distance pair
original_distance.updateMinElement();
std::pair<Size, Size> min = original_distance.getMinElementCoordinates();
Size overall_cluster_steps(original_distance.dimensionsize());
startProgress(0, original_distance.dimensionsize(), "clustering data");
while (original_distance(min.first, min.second) < threshold)
{
//grow the tree
cluster_tree.push_back(BinaryTreeNode(*(clusters[min.second].begin()), *(clusters[min.first].begin()), original_distance(min.first, min.second)));
if (cluster_tree.back().left_child > cluster_tree.back().right_child)
{
std::swap(cluster_tree.back().left_child, cluster_tree.back().right_child);
}
if (original_distance.dimensionsize() > 2)
{
//pick minimum-distance pair i,j and merge them
//pushback elements of second to first (and then erase second)
clusters[min.second].insert(clusters[min.first].begin(), clusters[min.first].end());
// erase first one
clusters.erase(clusters.begin() + min.first);
//update original_distance matrix
//complete linkage: new distance between clusters is the minimum distance between elements of each cluster
//lance-williams update for d((i,j),k): 0.5* d(i,k) + 0.5* d(j,k) + 0.5* |d(i,k)-d(j,k)|
for (Size k = 0; k < min.second; ++k)
{
float dik = original_distance.getValue(min.first, k);
float djk = original_distance.getValue(min.second, k);
original_distance.setValueQuick(min.second, k, (0.5f * dik + 0.5f * djk + 0.5f * std::fabs(dik - djk)));
}
for (Size k = min.second + 1; k < original_distance.dimensionsize(); ++k)
{
float dik = original_distance.getValue(min.first, k);
float djk = original_distance.getValue(min.second, k);
original_distance.setValueQuick(k, min.second, (0.5f * dik + 0.5f * djk + 0.5f * std::fabs(dik - djk)));
}
//reduce
original_distance.reduce(min.first);
//update minimum-distance pair
original_distance.updateMinElement();
//get new min-pair
min = original_distance.getMinElementCoordinates();
}
else
{
break;
}
setProgress(overall_cluster_steps - original_distance.dimensionsize());
//repeat until only two cluster remains or threshold exceeded, last step skips matrix operations
}
//fill tree with dummy nodes
Size sad(*clusters.front().begin());
for (Size i = 1; i < clusters.size() && (cluster_tree.size() < cluster_tree.capacity()); ++i)
{
cluster_tree.push_back(BinaryTreeNode(sad, *clusters[i].begin(), -1.0));
}
//~ while(cluster_tree.size() < cluster_tree.capacity())
//~ {
//~ cluster_tree.push_back(BinaryTreeNode(0,1,-1.0));
//~ }
endProgress();
}
void SingleLinkage::operator()(DistanceMatrix<float> & original_distance, std::vector<BinaryTreeNode> & cluster_tree, const float threshold /*=1*/) const
{
// input MUST have >= 2 elements!
if (original_distance.dimensionsize() < 2)
{
throw ClusterFunctor::InsufficientInput(__FILE__, __LINE__, OPENMS_PRETTY_FUNCTION, "Distance matrix to start from only contains one element");
}
cluster_tree.clear();
if (threshold < 1)
{
LOG_ERROR << "You tried to use Single Linkage clustering with a threshold. This is currently not supported!" << std::endl;
throw Exception::NotImplemented(__FILE__, __LINE__, OPENMS_PRETTY_FUNCTION);
}
//SLINK
std::vector<Size> pi;
pi.reserve(original_distance.dimensionsize());
std::vector<float> lambda;
lambda.reserve(original_distance.dimensionsize());
startProgress(0, original_distance.dimensionsize(), "clustering data");
//initialize first pointer values
pi.push_back(0);
lambda.push_back(std::numeric_limits<float>::max());
for (Size k = 1; k < original_distance.dimensionsize(); ++k)
{
std::vector<float> row_k;
row_k.reserve(k);
//initialize pointer values for element to cluster
pi.push_back(k);
lambda.push_back(std::numeric_limits<float>::max());
// get the right distances
for (Size i = 0; i < k; ++i)
{
row_k.push_back(original_distance.getValue(i, k));
}
//calculate pointer values for element k
for (Size i = 0; i < k; ++i)
{
if (lambda[i] >= row_k[i])
{
row_k[pi[i]] = std::min(row_k[pi[i]], lambda[i]);
lambda[i] = row_k[i];
pi[i] = k;
}
else
{
row_k[pi[i]] = std::min(row_k[pi[i]], row_k[i]);
}
}
//update clustering if necessary
for (Size i = 0; i < k; ++i)
{
if (lambda[i] >= lambda[pi[i]])
{
pi[i] = k;
}
}
setProgress(k);
}
for (Size i = 0; i < pi.size() - 1; ++i)
{
//strict order is always kept in algorithm: i < pi[i]
cluster_tree.push_back(BinaryTreeNode(i, pi[i], lambda[i]));
//~ std::cout << i << '\n' << pi[i] << '\n' << lambda[i] << std::endl;
}
//sort pre-tree
std::sort(cluster_tree.begin(), cluster_tree.end(), compareBinaryTreeNode);
// convert -pre-tree to correct format
for (Size i = 0; i < cluster_tree.size(); ++i)
{
if (cluster_tree[i].right_child < cluster_tree[i].left_child)
{
std::swap(cluster_tree[i].left_child, cluster_tree[i].right_child);
}
for (Size k = i + 1; k < cluster_tree.size(); ++k)
{
if (cluster_tree[k].left_child == cluster_tree[i].right_child)
{
cluster_tree[k].left_child = cluster_tree[i].left_child;
}
else if (cluster_tree[k].left_child > cluster_tree[i].right_child)
{
--cluster_tree[k].left_child;
}
if (cluster_tree[k].right_child == cluster_tree[i].right_child)
{
cluster_tree[k].right_child = cluster_tree[i].left_child;
}
else if (cluster_tree[k].right_child > cluster_tree[i].right_child)
{
--cluster_tree[k].right_child;
}
}
}
//~ prepare to redo clustering to get all indices for binarytree in min index element representation
std::vector<std::set<Size> > clusters(original_distance.dimensionsize());
for (Size i = 0; i < original_distance.dimensionsize(); ++i)
{
clusters[i].insert(i);
}
for (Size cluster_step = 0; cluster_step < cluster_tree.size(); ++cluster_step)
{
Size new_left_child = *(clusters[cluster_tree[cluster_step].left_child].begin());
Size new_right_child = *(clusters[cluster_tree[cluster_step].right_child].begin());
clusters[cluster_tree[cluster_step].left_child].insert(clusters[cluster_tree[cluster_step].right_child].begin(), clusters[cluster_tree[cluster_step].right_child].end());
clusters.erase(clusters.begin() + cluster_tree[cluster_step].right_child);
std::swap(cluster_tree[cluster_step].left_child, new_left_child);
std::swap(cluster_tree[cluster_step].right_child, new_right_child);
if (cluster_tree[cluster_step].left_child > cluster_tree[cluster_step].right_child)
{
std::swap(cluster_tree[cluster_step].left_child, cluster_tree[cluster_step].right_child);
}
}
endProgress();
}
