def test_sample_network_simulation(self):
sparsity = 1
noise_prob = 0
sim_partial_network = sim.sample_network(self.cluster_sizes,sparsity,noise_prob)
#without sparsity should completely recover network
#and no noise either
assert (self.sim_full_network != sim_partial_network).nnz == 0
def test_sample_network_noise(self):
#with complete noise should recover opposite of matrix
#could write another probabilistic test for the case of partial noise
#but this is OK
sparsity = 1
noise_prob = 1
sim_partial_network = sim.sample_network(self.cluster_sizes,sparsity,noise_prob)
#should recover opposite of matrix
assert (self.sim_full_network == sim_partial_network).nnz == 0
def test_sample_network_sparse_quantity(self):
#with sparsity should in expectation recover a lot of the network
sparsity = 0.5
noise_prob = 0
sim_partial_network = sim.sample_network(self.cluster_sizes,sparsity,noise_prob, symmetric=False)
expected_num_edges = self.network_size**2 * sparsity
#Probabilistic test: MAY FAIL WITH CORRECT BEHAVIOR (BUT W/ LOW PROBABILITY)
#by a Chernoff bound, this should deviate from expected number of edges
#by more than t*(network_size/2)
#with probability 2*e^(-t^2/2)
#here set t = 4, so this tests passes with probability 1 - 2*e^-8
actual_num_edges = sim_partial_network.nnz
print expected_num_edges, actual_num_edges
assert abs(actual_num_edges - expected_num_edges) <= 20
def clustering_pipeline(network_params, clustering_params):
#Create network
cluster_sizes, sparsity, noise_prob = network_params
num_clusters = len(cluster_sizes)
network = sim.sample_network(cluster_sizes, sparsity, noise_prob)
rows, cols = network.nonzero()
#Assign ground truth labels (the first "cluster_size" are in cluster 0,
#next are in cluster 1, etc.)
cluster_labels = list()
for cluster_index in range(len(cluster_sizes)):
cluster_labels += [cluster_index] * cluster_sizes[cluster_index]
cluster_labels = np.asarray(cluster_labels)
#perform clustering
cluster_sizes, method, completion_alg, completion_params, mode = clustering_params
cluster_predictions = cluster_signed_network(network, cluster_sizes, method, \
completion_alg, completion_params, mode)
cluster_accuracy = evaluate_cluster_accuracy(cluster_predictions, cluster_labels, \
rows, cols)
return cluster_accuracy
def run_experiment():
simulated = False
real = True
use_moi = True
use_hoc = True
use_svp = True
use_sgd_sh = False
use_sgd_sig = False
use_als = True
adj_matrix = None
if simulated:
cluster_sizes = [100, 200, 300, 400]
sparsity_level = 0.01175
noise_prob = 0
print "creating adjacency matrix..."
adj_matrix = sim.sample_network(cluster_sizes, sparsity_level,
noise_prob)
elif real:
data_file_name = "data/Preprocessed Data/small_network.npy"
#data_file_name = "data/Preprocessed Data/wiki_elections_csr.npy"
try:
adj_matrix = np.load(data_file_name).item()
except Exception as e:
raise ValueError("could not load adj matrix from file: ", e)
if use_moi:
print "performing MOI..."
max_cycle_order_moi = 10
discount = [0.5**i for i in range(3, max_cycle_order_moi + 1)]
#max_cycle_order_moi = np.inf
#discount = 0.0001
num_folds = 5
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
moi.kfoldcv_moi(adj_matrix, discount, max_cycle_order_moi, num_folds)
print "MOI results: "
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
if use_hoc:
print "performing HOC..."
max_cycle_order_hoc = 5
num_folds = 10
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
hoc.hoc_learning_pipeline(adj_matrix, max_cycle_order_hoc)
print "HOC results:"
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
alg = ""
alg_params = None
#settings if using SGD
if use_sgd_sh or use_sgd_sig:
#Parameters used for this experiment
#https://www.cs.uic.edu/~liub/KDD-cup-2007/proceedings/Regular-Paterek.pdf
learning_rate = 1000 #0.05 for square hinge
tol = adj_matrix.nnz / 10
max_iter = 20
reg_param = 10 #0.5 for square hinge
dim = 100
num_folds_mf = 10
#Bundle up these parameters and use this algorithm
if use_sgd_sh:
loss_type = "squarehinge" #"sigmoid"
alg_params = (learning_rate, loss_type, tol, max_iter, reg_param,
dim)
alg = "sgd"
print "performing SGD with square-hinge loss..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SGD_SH results:"
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
if use_sgd_sig:
loss_type = "sigmoid"
alg_params = (learning_rate, loss_type, tol, max_iter, reg_param,
dim)
alg = "sgd"
print "performing SGD with sigmoid loss..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SGD_SIG results:"
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
#settings if using als
if use_als:
#Parameters used for this experiment
max_iter = 2
dim = 40
#Bundle up these parameters and use this algorithm
alg_params = (max_iter, dim)
alg = "als"
num_folds_mf = 10
print "performing ALS..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "ALS results:"
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
#settings if using SVP
if use_svp:
#Parameters used for this experiment
rank = 40
tol = 100
max_iter = 5
step_size = 1
#Bundle up these parameters and use this algorithm
alg_params = (rank, tol, max_iter, step_size)
alg = "svp"
num_folds_mf = 10
print "performing SVP..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SVP results:"
print("Accuracy: average %f with standard error %f" %
(avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" %
(avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" %
(avg_time, stderr_time))
print
'''
cluster_sizes = [2,3,4]
sparsity_level = 0.5
noise_prob = 0
adj_matrix = sim.sample_network(cluster_sizes, sparsity_level, noise_prob)
#print adj_matrix.A
signed_laplacian(adj_matrix).A
'''
adj_matrix = None
if simulated:
cluster_sizes = [500,500]#[100,200,300,400]
sparsity_level = 0.5#0.01175
noise_prob = 0
print "creating adjacency matrix..."
adj_matrix = sim.sample_network(cluster_sizes, sparsity_level, noise_prob)
elif real:
#data_file_name = "Preprocessed Data/small_network.npy"
data_file_name = "../data/Preprocessed Data/wiki_elections_csr.npy"
try:
adj_matrix = np.load(data_file_name).item()
except Exception as e:
raise ValueError("could not load adj matrix from file: ", e)
avg_acc, avg_fpr = kfoldcv(adj_matrix, num_folds = 20)
print("Accuracy %f and false positive rate %f" % (avg_acc, avg_fpr))
def run_experiment():
simulated = False
real = True
use_moi = True
use_hoc = True
use_svp = True
use_sgd_sh = False
use_sgd_sig = False
use_als = True
adj_matrix = None
if simulated:
cluster_sizes = [100,200,300,400]
sparsity_level = 0.01175
noise_prob = 0
print "creating adjacency matrix..."
adj_matrix = sim.sample_network(cluster_sizes, sparsity_level, noise_prob)
elif real:
data_file_name = "data/Preprocessed Data/small_network.npy"
#data_file_name = "data/Preprocessed Data/wiki_elections_csr.npy"
try:
adj_matrix = np.load(data_file_name).item()
except Exception as e:
raise ValueError("could not load adj matrix from file: ", e)
if use_moi:
print "performing MOI..."
max_cycle_order_moi = 10
discount = [0.5**i for i in range(3, max_cycle_order_moi + 1)]
#max_cycle_order_moi = np.inf
#discount = 0.0001
num_folds = 5
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
moi.kfoldcv_moi(adj_matrix, discount, max_cycle_order_moi, num_folds)
print "MOI results: "
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print
if use_hoc:
print "performing HOC..."
max_cycle_order_hoc = 5
num_folds = 10
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
hoc.hoc_learning_pipeline(adj_matrix, max_cycle_order_hoc)
print "HOC results:"
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print
alg = ""
alg_params = None
#settings if using SGD
if use_sgd_sh or use_sgd_sig:
#Parameters used for this experiment
#https://www.cs.uic.edu/~liub/KDD-cup-2007/proceedings/Regular-Paterek.pdf
learning_rate = 1000#0.05 for square hinge
tol = adj_matrix.nnz/10
max_iter = 20
reg_param = 10#0.5 for square hinge
dim = 100
num_folds_mf = 10
#Bundle up these parameters and use this algorithm
if use_sgd_sh:
loss_type = "squarehinge" #"sigmoid"
alg_params = (learning_rate, loss_type, tol, max_iter, reg_param, dim)
alg = "sgd"
print "performing SGD with square-hinge loss..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SGD_SH results:"
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print
if use_sgd_sig:
loss_type = "sigmoid"
alg_params = (learning_rate, loss_type, tol, max_iter, reg_param, dim)
alg = "sgd"
print "performing SGD with sigmoid loss..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SGD_SIG results:"
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print
#settings if using als
if use_als:
#Parameters used for this experiment
max_iter = 2
dim = 40
#Bundle up these parameters and use this algorithm
alg_params = (max_iter, dim)
alg = "als"
num_folds_mf = 10
print "performing ALS..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "ALS results:"
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print
#settings if using SVP
if use_svp:
#Parameters used for this experiment
rank = 40
tol = 100
max_iter = 5
step_size = 1
#Bundle up these parameters and use this algorithm
alg_params = (rank, tol, max_iter, step_size)
alg = "svp"
num_folds_mf = 10
print "performing SVP..."
avg_acc, stderr_acc, avg_fpr, stderr_fpr, avg_time, stderr_time = \
mf.kfold_CV_pipeline(adj_matrix, alg, alg_params, num_folds_mf)
print "SVP results:"
print("Accuracy: average %f with standard error %f" % (avg_acc, stderr_acc))
print("False positive rate: average %f with standard error %f" % (avg_fpr, stderr_fpr))
print("Model running time: average %f with standard error %f" % (avg_time, stderr_time))
print