def categorize(self, element):
dists = {}
for number_id in self.data:
dists[number_id] = Utils.euclidean_distance( self.data[number_id], element )
s_dists = list( dists.values() )
s_dists.sort()
k_th = s_dists[self.k-1]
k_labels = {}
for number_id in dists:
if dists[number_id] <= k_th:
label = self.labels[number_id]
if label in k_labels:
k_labels[label] += 1
else:
k_labels[label] = 1
return max(k_labels, key=k_labels.get)
def measure(self):
x = []
y = []
for i in range(len(self.data)):
k = i + 1
knn = KNN(self.labels, self.data, k, 'regression')
KFoldCV_error = knn.KFoldCV(5)
if KFoldCV_error > -1:
x.append(1.0 / k)
y.append(KFoldCV_error)
Utils.plot2D(x, y)
def measure(self): x = [] y = [] for i in range( len(self.data) ): k = i+1 knn = KNN(self.labels, self.data, k, 'regression') KFoldCV_error = knn.KFoldCV( 5 ) if KFoldCV_error > -1: x.append( 1.0 / k ) y.append( KFoldCV_error ) Utils.plot2D(x,y)
def absolute_error(self, clusters, means): ae = 0.0 for cluster in clusters: sum_dissimilarities = 0 for element in clusters[cluster]: sum_dissimilarities += Utils.euclidean_distance( element, means[cluster] ) ae += sum_dissimilarities return ae
def absolute_error(self, clusters, means):
ae = 0.0
for cluster in clusters:
sum_dissimilarities = 0
for element in clusters[cluster]:
sum_dissimilarities += Utils.euclidean_distance(
element, means[cluster])
ae += sum_dissimilarities
return ae
def MSE(self, clusters, means): mse = 0.0 for cluster in clusters: mse_c = 0.0 for element in clusters[cluster]: mse_c += pow( Utils.euclidean_distance( element, means[cluster] ), 2 ) if len(clusters[cluster]) > 0: mse_c = ( mse_c / len(clusters[cluster]) ) mse += mse_c return (mse / len(clusters) )
def MSE(self, clusters, means):
mse = 0.0
for cluster in clusters:
mse_c = 0.0
for element in clusters[cluster]:
mse_c += pow(Utils.euclidean_distance(element, means[cluster]),
2)
if len(clusters[cluster]) > 0:
mse_c = (mse_c / len(clusters[cluster]))
mse += mse_c
return (mse / len(clusters))
def getMedoid(self, cluster_elements):
if len(cluster_elements) == 0:
return self.empty_row
N = len(cluster_elements[0])
centroid = self.getCentroid(cluster_elements)
# just a big number to start
min_dist = self.big_number
medoid = self.empty_row
for element in cluster_elements:
dist = Utils.euclidean_distance(centroid, element)
if dist < min_dist:
min_dist = dist
medoid = element
return medoid
def getMedoid(self, cluster_elements): if len(cluster_elements) == 0: return self.empty_row N = len(cluster_elements[0]) centroid = self.getCentroid( cluster_elements ) # just a big number to start min_dist = self.big_number medoid = self.empty_row for element in cluster_elements: dist = Utils.euclidean_distance( centroid, element ) if dist < min_dist: min_dist = dist medoid = element return medoid
def KMeansCore(self, data, k, mean_method, means):
K = {}
min_AE = self.big_number
prev_AE = 0
min_means = {}
min_K = {}
iter_flag = True
# do iterative relocation, until nothing changes
while (iter_flag):
for cluster in range(k):
K[cluster] = []
# assign each element to the cluster which has the closest mean
for element in data:
# for start, just set a big max number, and just pick the cluster 0 as the closest cluster
min_dist = self.big_number
closest_cluster = 0
for cluster in range(k):
dist = Utils.euclidean_distance(means[cluster], element)
if dist < min_dist:
min_dist = dist
closest_cluster = cluster
K[closest_cluster].append(element)
means[closest_cluster] = element
# calculate new mean for each cluster
if mean_method == self.mean_methods[0]:
for cluster in range(k):
means[cluster] = self.getCentroid(K[cluster])
elif mean_method == self.mean_methods[1]:
for cluster in range(k):
means[cluster] = self.getMedoid(K[cluster])
# calculate the absolute error
if mean_method == self.mean_methods[0]:
AE = self.MSE(K, means)
elif mean_method == self.mean_methods[1]:
AE = self.absolute_error(K, means)
# keep the clustering that minimizes the absolute error
if AE < min_AE:
min_AE = AE
min_means = means
min_K = K
# stop when nothing changes
if prev_AE == AE:
iter_flag = False
else:
prev_AE = AE
return (min_K, min_means, min_AE)
def KMeansCore(self, data, k, mean_method, means):
K = {}
min_AE = self.big_number
prev_AE = 0
min_means = {}
min_K = {}
iter_flag = True
# do iterative relocation, until nothing changes
while(iter_flag):
for cluster in range(k):
K[cluster] = []
# assign each element to the cluster which has the closest mean
for element in data:
# for start, just set a big max number, and just pick the cluster 0 as the closest cluster
min_dist = self.big_number
closest_cluster = 0
for cluster in range(k):
dist = Utils.euclidean_distance( means[cluster], element )
if dist < min_dist:
min_dist = dist
closest_cluster = cluster
K[closest_cluster].append( element )
means[closest_cluster] = element
# calculate new mean for each cluster
if mean_method == self.mean_methods[0]:
for cluster in range(k):
means[cluster] = self.getCentroid( K[cluster] )
elif mean_method == self.mean_methods[1]:
for cluster in range(k):
means[cluster] = self.getMedoid( K[cluster] )
# calculate the absolute error
if mean_method == self.mean_methods[0]:
AE = self.MSE(K, means)
elif mean_method == self.mean_methods[1]:
AE = self.absolute_error(K, means)
# keep the clustering that minimizes the absolute error
if AE < min_AE:
min_AE = AE
min_means = means
min_K = K
# stop when nothing changes
if prev_AE == AE:
iter_flag = False
else:
prev_AE = AE
return (min_K, min_means, min_AE)
