def test_select_k_nearest_neighbors_numpy_matrix():
matrix = numpy.array([(0,), (1,), (2,)])
center = numpy.array([0])
assert set(knn.select_k_nearest_neighbors(matrix, center, 2)) == set([0, 1])
assert set(knn.select_k_nearest_neighbors(matrix, center, 1)) == set([0])
matrix = numpy.array([(0, 0), (1, 1), (2, 2)])
center = numpy.array([0, 0])
assert set(knn.select_k_nearest_neighbors(matrix, center, 2)) == set([0, 1])
assert set(knn.select_k_nearest_neighbors(matrix, center, 3)) == set([0, 1, 2])
def test_select_k_nearest_neighbors():
matrix = [(0,), (1,), (2,)]
center = [0]
assert set(knn.select_k_nearest_neighbors(matrix, center, 2)) == set([0, 1])
assert set(knn.select_k_nearest_neighbors(matrix, center, 1)) == set([0])
matrix = [(0, 0), (1, 1), (2, 2)]
center = [0, 0]
assert set(knn.select_k_nearest_neighbors(matrix, center, 2)) == set([0, 1])
assert set(knn.select_k_nearest_neighbors(matrix, center, 3)) == set([0, 1, 2])
def clean_dataset_depuration(input_matrix, target_matrix, k=3, k_prime=2):
"""Clean a dataset with the Depuration procedure.
See section 3.1 of "Analysis of new techniques to obtain quality training sets".
"""
if not isinstance(input_matrix, numpy.ndarray):
input_matrix = numpy.array(input_matrix)
if not isinstance(target_matrix, numpy.ndarray):
target_matrix = numpy.array(target_matrix)
patterns = zip(input_matrix, target_matrix)
indices = range(len(input_matrix))
if not ((k + 1) / 2 <= k_prime and k_prime <= k):
raise ValueError('k_prime must be between (k + 1) / 2 and k')
kept_inputs = []
kept_targets = []
changed_patterns = []
removed_patterns = []
for i in indices:
# Find k-NN of patternIi in patterns - {pattern_i}
# We do this by finding k+1 nearest indices, and ignoring index i
k_nearest = knn.select_k_nearest_neighbors(input_matrix,
input_matrix[i], k + 1)
k_nearest.remove(i)
# if a class has at least k_prime representatives
# among the k neighbours
class_counts = _count_classes(target_matrix[k_nearest])
removed = True
for class_, count in class_counts.iteritems():
if count >= k_prime:
# Change the label of pattern to that class
# and add pattern to cleaned_patterns
# Keep input
# and the common class
kept_inputs.append(input_matrix[i])
kept_targets.append(class_)
# Check if new pattern is different than old pattern, track changed
if class_ != tuple(target_matrix[i]):
changed_patterns.append(i)
removed = False
break
if removed:
# discard pattern (do not add to kept lists)
removed_patterns.append(i)
return ((numpy.array(kept_inputs), numpy.array(kept_targets)),
changed_patterns, removed_patterns)