def get_augmented_gower_selection(D):
"""
Do a spectral sign split with neighbor joining fallback.
The first choice is to return indices corresponding to
positive elements of the dominant eigenvector of the gower matrix.
If this defines a degenerate bipartition,
then neighbor joining is used as a fallback.
@param D: a distance matrix
@return: the set of selected indices
"""
n = len(D)
if n < 4:
raise ValueError('expected a distance matrix with at least four rows')
# get the gower matrix
G = MatrixUtil.double_centered(numpy.array(D))
# get the dominant eigenvector
eigenvalues, eigenvector_transposes = linalg.eigh(G)
eigenvectors = eigenvector_transposes.T
dominant_value, dominant_vector = max(
(abs(w), v) for w, v in zip(eigenvalues, eigenvectors))
# get the bipartition defined by the dominant eigenvector
selection = set(i for i, x in enumerate(dominant_vector) if x > 0)
complement = set(range(n)) - selection
# if the bipartition is degenerate then resort to neighbor joining
if min(len(selection), len(complement)) < 2:
selection = set(NeighborJoining.get_neighbors(D))
return selection
def get_augmented_gower_selection(D):
"""
Do a spectral sign split with neighbor joining fallback.
The first choice is to return indices corresponding to
positive elements of the dominant eigenvector of the gower matrix.
If this defines a degenerate bipartition,
then neighbor joining is used as a fallback.
@param D: a distance matrix
@return: the set of selected indices
"""
n = len(D)
if n < 4:
raise ValueError('expected a distance matrix with at least four rows')
# get the gower matrix
G = MatrixUtil.double_centered(numpy.array(D))
# get the dominant eigenvector
eigenvalues, eigenvector_transposes = linalg.eigh(G)
eigenvectors = eigenvector_transposes.T
dominant_value, dominant_vector = max((abs(w), v) for w, v in zip(eigenvalues, eigenvectors))
# get the bipartition defined by the dominant eigenvector
selection = set(i for i, x in enumerate(dominant_vector) if x > 0)
complement = set(range(n)) - selection
# if the bipartition is degenerate then resort to neighbor joining
if min(len(selection), len(complement)) < 2:
selection = set(NeighborJoining.get_neighbors(D))
return selection
def _get_any_selection(self, distance_matrix):
"""
@param distance_matrix: a numpy or row major distance matrix
@return: a set of selected indices representing one of the two parts of the bipartition
"""
return set(NeighborJoining.get_neighbors(distance_matrix))
