def __init__(self, G, G_test, root="results/lp/"):
super(LinkPredictionTuning, self).__init__(G, root=root)
self.task = "lp"
train_E = G.edges
train_E_false = self.GetNegativeEdges(G, len(train_E))
test_E = G_test.edges
test_E_false = self.GetNegativeEdges(G_test, len(test_E))
self.split = EvalSplit()
self.split.set_splits(train_E,
train_E_false=train_E_false,
test_E=test_E,
test_E_false=test_E_false,
TG=G)
self.training_graph = create_self_defined_dataset(root_dir="",
name_dict={},
name="training " +
self.tipo,
weighted=True,
directed=False,
attributed=True)()
self.training_graph.set_g(G)
self.evaluator = LPEvaluator(self.split)
def main():
# Initialize some parameters
inpath = list()
nw_names = ['network',
'blogCatalog'] # Stores the names of the networks evaluated
inpath.append("../evalne/tests/data/network.edgelist")
# inpath.append("../../data/BlogCatalog/blog.edgelist")
outpath = "./output/"
if not os.path.exists(outpath):
os.makedirs(outpath)
directed = False # indicates if the graphs are directed or undirected
delimiters = (',', '\t') # indicates the delimiter in the original graph
repeats = 2 # number of time the experiment will be repeated
# Create a scoresheet to store the results
scoresheet = Scoresheet(tr_te='test')
for i in range(len(inpath)):
# Create folders for the evaluation results (one per input network)
if not os.path.exists(outpath):
os.makedirs(outpath)
# Load and preprocess the graph
G = preprocess(inpath[i], outpath, delimiters[i], directed)
# For each repeat of the experiment generate new data splits
for repeat in range(repeats):
print('Repetition {} of experiment'.format(repeat))
# Generate one train/test split with default parameters
traintest_split = EvalSplit()
traintest_split.compute_splits(G,
nw_name=nw_names[i],
train_frac=0.8,
split_id=repeat)
trainvalid_split = EvalSplit()
trainvalid_split.compute_splits(traintest_split.TG,
nw_name=nw_names[i],
train_frac=0.9,
split_id=repeat)
# Create an evaluator
nee = LPEvaluator(traintest_split, trainvalid_split)
# Evaluate baselines
eval_baselines(nee, directed, scoresheet)
# Evaluate other NE methods
eval_other(nee, scoresheet)
# Write results averaged over exp repeats to a single file
scoresheet.write_tabular(filename=os.path.join(outpath, 'eval_output.txt'),
metric='auroc')
print("End of evaluation")
def main(args=None):
cpu_number = multiprocessing.cpu_count()
parser = argparse.ArgumentParser(description='Path of networks')
parser.add_argument('-n', type=str, help='Multiplex 1')
parser.add_argument('-m', type=str, help='Multiplex 2')
parser.add_argument('-b', type=str, help='Bipartite')
args = parser.parse_args(args)
print(args)
########################################################################
# Parameters multiverse and train/test
########################################################################
EMBED_DIMENSION = 128
CLOSEST_NODES = np.int64(300)
NUM_SAMPLED = np.int64(10)
LEARNING_RATE = np.float64(0.01)
KL = False
NB_CHUNK = np.int64(1)
CHUNK_SIZE = np.int64(100)
NUM_STEPS_1 = np.int64(100 * 10**6 / CHUNK_SIZE)
# If toy example
#EMBED_DIMENSION = 128
#CLOSEST_NODES = np.int64(2)
#NUM_SAMPLED = np.int64(10)
#LEARNING_RATE = np.float64(0.01)
#KL = False
#NB_CHUNK = np.int64(1)
#CHUNK_SIZE = np.int64(2)
#NUM_STEPS_1 = np.int64(100*10**6/CHUNK_SIZE)
train_frac = 0.7
solver = 'lbfgs'
max_iter = 1000
split_alg = 'random'
lp_model = RandomForestClassifier(n_estimators=400, criterion='gini', max_depth=None, min_samples_split=2, min_samples_leaf=1, min_weight_fraction_leaf=0.0, \
max_features='auto', max_leaf_nodes=None, min_impurity_decrease=0.0, min_impurity_split=None, bootstrap=True,\
oob_score=True, n_jobs=cpu_number, random_state=777, verbose=0, warm_start=False)
graph_name = 'Test_Eval'
##################################################################################
# !! Careful !!
# Check if nodes in the bipartite have the same nodes in the multiplex
# networks. If not you have to remove the nodes in the multiplexes not included in the
# bipartites
##################################################################################
###################################################################################
# EvalNE Link prediction processing
###################################################################################
data_bipartite = pd.read_csv(args.b, delimiter=' ', header=None)
data_bipartite = data_bipartite.drop(columns=[0, 3])
data_bipartite.to_csv('bipartite_2colformat.csv',
header=None,
index=None,
sep=' ')
G_hetereogeneous = f.preprocess('bipartite_2colformat.csv', '.', ' ',
False, False, True)
print('Preprocessing done')
G_hetereogeneous_traintest_split = EvalSplit()
G_hetereogeneous_traintest_split.compute_splits(G_hetereogeneous,
split_alg=split_alg,
train_frac=train_frac,
owa=False)
nee = LPEvaluator(G_hetereogeneous_traintest_split,
dim=EMBED_DIMENSION,
lp_model=lp_model)
G_heterogeneous_split = (G_hetereogeneous_traintest_split.TG)
os.replace('bipartite_2colformat.csv',
'./Generated_graphs/' + 'bipartite_2colformat.csv')
print('Splitting done')
# Write the bipartite training graph for multiverse in extended edgelist format 'layer n1 n2 weight'
file_multi = open('bipartite_training_graph_' + '_' + graph_name, 'w+')
tmp_array_het = []
tmp_array_het = np.asarray(G_heterogeneous_split.edges)
for i in range(len(tmp_array_het[:, 0])):
if tmp_array_het[i, 0] in list(data_bipartite[2]):
tmp = tmp_array_het[i, 0]
tmp_array_het[i, 0] = tmp_array_het[i, 1]
tmp_array_het[i, 1] = tmp
tmp_array_het = np.hstack((tmp_array_het, np.ones(
(len(tmp_array_het), 1))))
tmp_array_het = np.hstack((np.ones(
(len(tmp_array_het), 1)), tmp_array_het))
tmp_array_het = np.vstack(tmp_array_het)
tmp_array_het = np.int_(tmp_array_het)
np.savetxt(file_multi,
tmp_array_het,
fmt='%s',
delimiter=' ',
newline=os.linesep)
file_multi.close()
os.replace(
'bipartite_training_graph_' + '_' + graph_name, './Generated_graphs/' +
'bipartite_training_graph_' + '_' + graph_name + '.txt')
###################################################################################
# MULTIVERSE
###################################################################################
r_readRDS = robjects.r['readRDS']
print('RWR-MH')
proc = subprocess.Popen(['Rscript', './RWR/GenerateSimMatrix_MH.R', \
'-n', '.' + args.n, \
'-m', '.' + args.m, \
'-b', '../Generated_graphs/'+ 'bipartite_training_graph_' + '_'+ graph_name+'.txt',
'-o', '../ResultsRWR/MatrixSimilarityMultiplexHet'+graph_name, '-c', str(cpu_number)])
proc.wait()
proc.kill()
print('RWR done')
r_DistancematrixPPI = r_readRDS(
'./ResultsRWR/MatrixSimilarityMultiplexHet' + graph_name + '.rds')
import gc
gc.collect()
########################################################################
# Processing of the network
########################################################################
reverse_data_DistancematrixPPI, list_neighbours, nodes, data_DistancematrixPPI, neighborhood, nodesstr \
= f.netpreprocess_hetero(r_DistancematrixPPI, CLOSEST_NODES)
########################################################################
# Initialization
########################################################################
embeddings = np.random.normal(0, 1, [np.size(nodes), EMBED_DIMENSION])
########################################################################
# Training and saving best embeddings
########################################################################
# Train and test during training
neighborhood = np.asarray(neighborhood)
nodes = np.asarray(nodes)
embeddings = f.train(neighborhood, nodes, list_neighbours, NUM_STEPS_1, NUM_SAMPLED, LEARNING_RATE, \
CLOSEST_NODES, CHUNK_SIZE, NB_CHUNK, embeddings, reverse_data_DistancematrixPPI)
X = dict(zip(range(embeddings.shape[0]), embeddings))
X = {str(int(nodesstr[key]) + 1): X[key] for key in X}
np.save('embeddings_MH', X)
date = datetime.datetime.now()
os.replace('embeddings_MH.npy',
'./ResultsMultiVERSE/' + 'embeddings_MH.npy')
########################################################################
# Link prediction for evaluation of MH
########################################################################
edge_emb = ['hadamard', 'weighted_l1', 'weighted_l2', 'average', 'cosine']
results_embeddings_methods = dict()
for i in range(len(edge_emb)):
tmp_result_multiverse = nee.evaluate_ne(data_split=nee.traintest_split,
X=X,
method="Multiverse",
edge_embed_method=edge_emb[i],
label_binarizer=lp_model)
results_embeddings_methods[tmp_result_multiverse.method + '_' + str(
edge_emb[i])] = tmp_result_multiverse.get_all()[1][4]
########################################################################
# Analysis and saving of the results
########################################################################
Result_file = 'Result_LinkpredMultiplexHet_' + graph_name + '_' + str(
date) + '.txt'
with open(Result_file, "w+") as overall_result:
print("%s: \n\
EMBED_DIMENSION: %s \n\
CLOSEST_NODES: %s \n\
NUM_STEPS_1: %s \n\
NUM_SAMPLED: %s \n\
LEARNING_RATE: %s \n\
CHUNK_SIZE: %s \n\
NB_CHUNK: %s \n\
train_frac: %s \n\
solver: %s \n\
max_iter: %s \n\
split_alg: %s \n\
" % (str(date), EMBED_DIMENSION, CLOSEST_NODES, NUM_STEPS_1,
NUM_SAMPLED, LEARNING_RATE, CHUNK_SIZE, NB_CHUNK,
train_frac, solver, max_iter, split_alg),
file=overall_result)
print('Overall MULTIVERSE AUC hadamard:',
results_embeddings_methods['Multiverse_hadamard'],
file=overall_result)
print('Overall MULTIVERSE AUC weighted_l1:',
results_embeddings_methods['Multiverse_weighted_l1'],
file=overall_result)
print('Overall MULTIVERSE AUC weighted_l2:',
results_embeddings_methods['Multiverse_weighted_l2'],
file=overall_result)
print('Overall MULTIVERSE AUC average:',
results_embeddings_methods['Multiverse_average'],
file=overall_result)
print('Overall MULTIVERSE AUC cosine:',
results_embeddings_methods['Multiverse_cosine'],
file=overall_result)
overall_result.close()
os.replace(Result_file, './ResultsMultiVERSE/' + Result_file)
print('End')
class LinkPredictionTuning(Tuning):
r"""
Clase general de entrenamiento y testeo de embeddings de grafos para la tarea de prediccion de enlaces.
Parameters
----------
G: NetworkX graph
Grafo de entrenamiento.
G_test: NetworkX graph
Grafo de testeo.
root: str
directorio en el que se guardaran los resultados
"""
def __init__(self, G, G_test, root="results/lp/"):
super(LinkPredictionTuning, self).__init__(G, root=root)
self.task = "lp"
train_E = G.edges
train_E_false = self.GetNegativeEdges(G, len(train_E))
test_E = G_test.edges
test_E_false = self.GetNegativeEdges(G_test, len(test_E))
self.split = EvalSplit()
self.split.set_splits(train_E,
train_E_false=train_E_false,
test_E=test_E,
test_E_false=test_E_false,
TG=G)
self.training_graph = create_self_defined_dataset(root_dir="",
name_dict={},
name="training " +
self.tipo,
weighted=True,
directed=False,
attributed=True)()
self.training_graph.set_g(G)
self.evaluator = LPEvaluator(self.split)
def GetNegativeEdges(self, G, n):
r"""
Metodo auxiliar que muestrea enlaces negativos.
Parameters
----------
G: NetworkX graph
Grafo bipartito.
n: int
cantidad de enlaces que muestrear.
"""
prop_nodes = [n for n, d in G.nodes(data=True) if d['bipartite'] == 0]
user_nodes = [n for n, d in G.nodes(data=True) if d['bipartite'] == 1]
non_edges = []
while len(non_edges) <= n:
random_prop = random.choice(prop_nodes)
random_user = random.choice(user_nodes)
edge = (random_prop, random_user)
if G.has_edge(*edge):
continue
else:
non_edges.append(edge)
return non_edges
def TestModel(self, emb, time=-1, method_name="method_name"):
r"""
Testea un embedding y lo guarda en el scoresheet.
Parameters
----------
emb: dict
diccionario de embeddings, llaves son los nodos y los valores una lista con el embedding
time: float
tiempo de ejecucion del metodo, para guardar en el scoresheet
method_name: str
nombre del metodo con el que guardar.
"""
df = pd.DataFrame(emb).T
X = df.T.to_dict("list")
X = {str(k): np.array(v)
for k, v in X.items()
} # tiene que ser array por que se hacen sumas
self.evaluator.dim = df.shape[1]
reslp = []
for edge_method in [
"weighted_l1", "weighted_l2", "hadamard", "average"
]:
#TO DO que no evalue en los 4 embeddings de enlaces
res = self.evaluator.evaluate_ne(self.split,
X=X,
method=method_name,
edge_embed_method=edge_method,
params={"nw_name": "GPI"})
res.params.update({'eval_time': time})
reslp.append(res)
self.scoresheet.log_results(reslp)
return reslp
def main(args=None):
cpu_number = multiprocessing.cpu_count()
parser = argparse.ArgumentParser(description='Path of networks')
parser.add_argument('-m', type=str, help='Multiplex')
args = parser.parse_args(args)
graph_path = args.m
########################################################################
# Parameters multiverse and train/test
########################################################################
EMBED_DIMENSION = 128
CLOSEST_NODES = np.int64(20)
NUM_SAMPLED = np.int64(3)
LEARNING_RATE = np.float64(0.01)
NB_CHUNK = np.int64(1)
CHUNK_SIZE = np.int64(10)
NUM_STEPS_1 = np.int64(100 * 10**6 / CHUNK_SIZE)
graph_name = os.path.basename(graph_path)
train_frac = 0.7
solver = 'lbfgs'
max_iter = 2000
split_alg = 'spanning_tree'
lp_model = LogisticRegressionCV(Cs=10, cv= 5, class_weight=None, dual=False, fit_intercept=True, intercept_scaling=1.0, max_iter=max_iter, \
multi_class='ovr', n_jobs=cpu_number, random_state=None, refit=True, scoring='roc_auc', solver=solver, tol=0.0001, verbose=0)
edge_data_by_type, _, all_nodes = f.load_network_data(graph_path)
nb_layers = len(edge_data_by_type.keys())
# Divide multiplex graph in several in edgelist format
for layer in range(nb_layers - 1):
file = open(
'multiplex_graph_layer_' + str(layer + 1) + '_' + graph_name, 'w+')
tmp_array = np.asarray(edge_data_by_type[str(layer + 1)])
np.savetxt(file, tmp_array, fmt='%s')
file.close()
os.replace(
'multiplex_graph_layer_' + str(layer + 1) + '_' + graph_name,
'Generated_graphs/' + 'multiplex_graph_layer_' + str(layer + 1) +
'_' + graph_name)
# Load each graph with EvalNE, preprocess and split train/test edges
nee = list()
G_original = list()
Gsplit = list()
traintestsplit = list()
for layer in range(nb_layers - 1):
G_original.append(
f.preprocess('./Generated_graphs/' + 'multiplex_graph_layer_' +
str(layer + 1) + '_' + graph_name,
'.',
' ',
directed=False,
relabel=False,
del_self_loops=True))
G_original_traintest_split = EvalSplit()
G_original_traintest_split.compute_splits(G_original[layer],
split_alg=split_alg,
train_frac=train_frac,
owa=False)
traintestsplit.append(G_original_traintest_split)
nee.append(
LPEvaluator(G_original_traintest_split,
dim=EMBED_DIMENSION,
lp_model=lp_model))
Gsplit.append(G_original_traintest_split.TG)
# Write the multiplex training graph for multiverse in extended edgelist format 'layer n1 n2 weight'
file_multi = open('multiverse_graph_' + 'training' + '_' + graph_name,
'w+')
matrix_train_edges = []
sorted_matrix_train_edges = []
tmp_array_multi = []
tmp_array_collapsed = []
for layer in range(nb_layers - 1):
tmp_array = np.asarray(Gsplit[layer].edges)
tmp_array = np.hstack((tmp_array, np.ones((len(tmp_array), 1))))
tmp_array = np.hstack(((layer + 1) * np.ones(
(len(tmp_array), 1)), tmp_array))
tmp_array = np.vstack(tmp_array)
tmp_array_multi.append(tmp_array)
tmp_array_mat_train_edges = np.asarray(Gsplit[layer].edges)
tmp_array_mat_train_edges = np.hstack(
(tmp_array_mat_train_edges,
np.ones((len(tmp_array_mat_train_edges), 1))))
tmp_array_mat_train_edges = np.hstack(((layer) * np.ones(
(len(tmp_array), 1)), tmp_array_mat_train_edges))
matrix_train_edges.append(tmp_array_mat_train_edges)
matrix_train_edges = sorted(tmp_array_mat_train_edges,
key=itemgetter(1))
sorted_matrix_train_edges.extend(matrix_train_edges)
matrix_train_edges = []
tmp_array_multi = np.vstack(tmp_array_multi)
tmp_array_multi = np.int_(tmp_array_multi)
np.savetxt(file_multi,
tmp_array_multi,
fmt='%s',
delimiter=' ',
newline=os.linesep)
file_multi.close()
os.replace(
'multiverse_graph_' + 'training' + '_' + graph_name,
'./Generated_graphs/' + 'multiverse_graph_' + 'training' + '_' +
graph_name + '.txt')
###################################################################################"
# MULTIVERSE
###################################################################################"
r_readRDS = robjects.r['readRDS']
proc = subprocess.Popen(['Rscript', './RWR/GenerateSimMatrix.R', \
'-n', '../Generated_graphs/'+'multiverse_graph_' + 'training' + '_'+ graph_name+'.txt', '-o', \
'../ResultsRWR/MatrixSimilarityMultiplex'+graph_name, '-c', str(cpu_number)])
proc.wait()
pid = proc.pid
proc.kill()
print('RWR done')
r_DistancematrixPPI = r_readRDS('./ResultsRWR/MatrixSimilarityMultiplex' +
graph_name + '.rds')
########################################################################
# Processing of the network
########################################################################
reverse_data_DistancematrixPPI, list_neighbours, nodes, data_DistancematrixPPI, nodes_incomponent, neighborhood, nodesstr \
= f.netpreprocess(r_DistancematrixPPI, graph_path, CLOSEST_NODES)
########################################################################
# Initialization
########################################################################
embeddings = np.random.normal(0, 1, [np.size(nodes), EMBED_DIMENSION])
########################################################################
# Training and saving best embeddings
########################################################################
nodes = np.asarray(nodes)
embeddings = f.train(neighborhood, nodes, list_neighbours, NUM_STEPS_1, NUM_SAMPLED, LEARNING_RATE, \
CLOSEST_NODES, CHUNK_SIZE, NB_CHUNK, embeddings, reverse_data_DistancematrixPPI)
X = dict(zip(range(embeddings.shape[0]), embeddings))
X = {str(int(nodesstr[key]) + 1): X[key] for key in X}
np.save('embeddings_M', X)
date = datetime.datetime.now()
os.replace('embeddings_M.npy', './ResultsMultiVERSE/' + 'embeddings_M.npy')
print('Embedding done')
########################################################################
# Evaluation on link prediction
########################################################################
edge_emb = ['hadamard', 'weighted_l1', 'weighted_l2', 'average', 'cosine']
results_embeddings_methods = dict()
date = datetime.datetime.now()
for layer in range(nb_layers - 1):
for i in range(len(edge_emb)):
tmp_result_multiverse = nee[layer].evaluate_ne(
data_split=nee[layer].traintest_split,
X=X,
method="Multiverse",
edge_embed_method=edge_emb[i],
label_binarizer=lp_model)
results_embeddings_methods[
tmp_result_multiverse.method + '_' + str(layer) +
str(edge_emb[i])] = tmp_result_multiverse.get_all()[1][4]
print('Evaluation done')
########################################################################
# Analysis and saving of the results
########################################################################
tmp_Multiverse_Result_hada = 0
tmp_Multiverse_Result_wl1 = 0
tmp_Multiverse_Result_wL2 = 0
tmp_Multiverse_Result_avg = 0
tmp_Multiverse_Result_cos = 0
for layer in range(nb_layers - 1):
tmp_Multiverse_Result_hada += results_embeddings_methods[
'Multiverse' + '_' + str(layer) + str(edge_emb[0])]
tmp_Multiverse_Result_wl1 += results_embeddings_methods[
'Multiverse' + '_' + str(layer) + str(edge_emb[1])]
tmp_Multiverse_Result_wL2 += results_embeddings_methods[
'Multiverse' + '_' + str(layer) + str(edge_emb[2])]
tmp_Multiverse_Result_avg += results_embeddings_methods[
'Multiverse' + '_' + str(layer) + str(edge_emb[3])]
tmp_Multiverse_Result_cos += results_embeddings_methods[
'Multiverse' + '_' + str(layer) + str(edge_emb[4])]
results_embeddings_methods[
'Multiverse_av_hadamard'] = tmp_Multiverse_Result_hada / (nb_layers -
1)
results_embeddings_methods[
'Multiverse_av_weighted_l1'] = tmp_Multiverse_Result_wl1 / (nb_layers -
1)
results_embeddings_methods[
'Multiverse_av_weighted_l2'] = tmp_Multiverse_Result_wL2 / (nb_layers -
1)
results_embeddings_methods[
'Multiverse_av_average'] = tmp_Multiverse_Result_avg / (nb_layers - 1)
results_embeddings_methods[
'Multiverse_av_cosine'] = tmp_Multiverse_Result_cos / (nb_layers - 1)
# Save results
Result_file = 'Result_Linkpred_Multiplex_' + graph_name + '_' + str(
date) + '.txt'
with open(Result_file, "w+") as overall_result:
print("%s: \n\
EMBED_DIMENSION: %s \n\
CLOSEST_NODES: %s \n\
NUM_STEPS_1: %s \n\
NUM_SAMPLED: %s \n\
LEARNING_RATE: %s \n\
CHUNK_SIZE: %s \n\
NB_CHUNK: %s \n\
train_frac: %s \n\
solver: %s \n\
max_iter: %s \n\
split_alg: %s \n\
" % (str(date), EMBED_DIMENSION, CLOSEST_NODES, NUM_STEPS_1,
NUM_SAMPLED, LEARNING_RATE, CHUNK_SIZE, NB_CHUNK,
train_frac, solver, max_iter, split_alg),
file=overall_result)
print('Overall MULTIVERSE AUC hadamard:',
results_embeddings_methods['Multiverse_av_hadamard'],
file=overall_result)
print('Overall MULTIVERSE AUC weighted_l1:',
results_embeddings_methods['Multiverse_av_weighted_l1'],
file=overall_result)
print('Overall MULTIVERSE AUC weighted_l2:',
results_embeddings_methods['Multiverse_av_weighted_l2'],
file=overall_result)
print('Overall MULTIVERSE AUC average:',
results_embeddings_methods['Multiverse_av_average'],
file=overall_result)
print('Overall MULTIVERSE AUC cosine:',
results_embeddings_methods['Multiverse_av_cosine'],
file=overall_result)
overall_result.close()
os.replace(Result_file, './ResultsMultiVERSE/' + Result_file)
print('End')
def evaluate(setup):
# Set the random seed
random.seed(setup.seed)
np.random.seed(setup.seed)
# Get input and output paths
inpaths = setup.inpaths
filename = '{}_eval_{}'.format(setup.task,
datetime.now().strftime("%m%d_%H%M"))
outpath = os.path.join(os.getcwd(), filename)
if not os.path.exists(outpath):
os.makedirs(outpath)
# Logging configuration (file opened in append mode)
logging.basicConfig(filename=os.path.join(outpath, 'eval.log'),
format='%(asctime)s - %(levelname)s: %(message)s',
datefmt='%d-%m-%y %H:%M:%S',
level=logging.INFO)
logging.info('Evaluation start')
if setup.task != 'nc':
logging.info('Running evaluation using classifier: {}'.format(
setup.lp_model))
# Create a Scoresheet object to store all results
if setup.task == 'nr':
scoresheet = Scoresheet(tr_te='train', precatk_vals=setup.precatk_vals)
else:
scoresheet = Scoresheet(tr_te='test', precatk_vals=setup.precatk_vals)
# Initialize some variables
edge_split_time = list()
lp_coef = dict()
repeats = setup.lp_num_edge_splits if setup.task == 'lp' else 1
t = tqdm(total=len(inpaths) * repeats)
t.set_description(desc='Progress on {} task'.format(setup.task))
# Loop over all input networks
for i in range(len(inpaths)):
logging.info('====== Evaluating {} network ======'.format(
setup.names[i]))
print('\nEvaluating {} network...'.format(setup.names[i]))
print('=====================================')
# Create path to store info per network if needed
nw_outpath = os.path.join(outpath, setup.names[i])
if setup.save_prep_nw or setup.curves != '':
if not os.path.exists(nw_outpath):
os.makedirs(nw_outpath)
# Load and preprocess the graph
G, ids = preprocess(setup, nw_outpath, i)
if setup.task == 'nc':
try:
labels = pp.read_labels(setup.labelpaths[i], idx_mapping=ids)
except (ValueError, IOError) as e:
logging.exception(
'Exception occurred while reading labels of `{}` network. Skipping network eval...'
.format(setup.names[i]))
break
# For each repeat of the experiment generate new edge splits
for repeat in range(repeats):
logging.info(
'------ Repetition {} of experiment ------'.format(repeat))
print('\nRepetition {} of experiment...'.format(repeat))
print('-------------------------------------')
# Create train and validation edge splits
traintest_split = EvalSplit()
trainvalid_split = EvalSplit()
split_time = time.time()
if setup.task == 'lp':
# For LP compute train/test and train/valid splits
traintest_split.compute_splits(G,
nw_name=setup.names[i],
train_frac=setup.traintest_frac,
split_alg=setup.split_alg,
owa=setup.owa,
fe_ratio=setup.fe_ratio,
split_id=repeat,
verbose=setup.verbose)
trainvalid_split.compute_splits(
traintest_split.TG,
nw_name=setup.names[i],
train_frac=setup.trainvalid_frac,
split_alg=setup.split_alg,
owa=setup.owa,
fe_ratio=setup.fe_ratio,
split_id=repeat,
verbose=setup.verbose)
# traintest_split.save_tr_graph(nw_outpath + '/TG_rep_{}'.format(repeat), ',', True, False, False)
# Create an LP evaluator
nee = LPEvaluator(traintest_split, trainvalid_split,
setup.embed_dim, setup.lp_model)
elif setup.task == 'nr':
# For NR set TG = G no train/valid split needed and get random subset of true and false edges for pred
pos_e, neg_e = stt.random_edge_sample(nx.adj_matrix(G),
setup.nr_edge_samp_frac,
nx.is_directed(G))
if len(pos_e) == 0:
logging.error(
'Sampling fraction {} on {} network returned 0 positive edges. Skipping evaluation...'
.format(setup.nr_edge_samp_frac, setup.names[i]))
break
traintest_split.set_splits(train_E=pos_e,
train_E_false=neg_e,
test_E=None,
test_E_false=None,
directed=nx.is_directed(G),
nw_name=setup.names[i],
TG=G)
# Create an NR evaluator
nee = NREvaluator(traintest_split, setup.embed_dim,
setup.lp_model)
else:
# Create an NC evaluator (train/valid fraction hardcoded to 10%)
nee = NCEvaluator(G, labels, setup.names[i],
setup.nc_num_node_splits,
setup.nc_node_fracs, 0.2, setup.embed_dim)
edge_split_time.append(time.time() - split_time)
# Evaluate baselines
if setup.lp_baselines is not None and setup.task != 'nc':
eval_baselines(setup, nee, i, scoresheet, repeat, nw_outpath)
# Evaluate other NE methods
if setup.methods_opne is not None or setup.methods_other is not None:
lp_coef = eval_other(setup, nee, i, scoresheet, repeat,
nw_outpath)
# Update progress bar
t.update(1)
# Store in a pickle file the results up to this point in evaluation
scoresheet.write_pickle(os.path.join(outpath, 'eval.pkl'))
# Store the results
if setup.scores is not None:
if setup.scores == 'all':
scoresheet.write_all(
filename=os.path.join(outpath, 'eval_output.txt'))
else:
scoresheet.write_tabular(filename=os.path.join(
outpath, 'eval_output.txt'),
metric=setup.scores)
scoresheet.write_tabular(filename=os.path.join(
outpath, 'eval_output.txt'),
metric='eval_time')
scoresheet.write_pickle(os.path.join(outpath, 'eval.pkl'))
# Close progress bar
t.close()
print('Average edge split times per dataset:')
print(setup.names)
print(np.array(edge_split_time).reshape(-1, repeats).mean(axis=1))
# if setup.task != 'nc':
# print('Coefficients of LP model ({}) for each NE method:'.format(setup.lp_model))
# print(lp_coef)
logging.info('Evaluation end\n\n')
from evalne.evaluation.evaluator import LPEvaluator
from evalne.evaluation.split import EvalSplit
from evalne.evaluation.score import Scoresheet
from evalne.utils import preprocess as pp
# Load and preprocess the network
#G = pp.load_graph('evalne/tests/data/network.edgelist')
G = pp.load_graph(
'../Graph_Conv_Neural_Nets/generic_datasets/Zachary-Karate/Zachary-Karate.edgelist'
)
G, _ = pp.prep_graph(G)
# Create an evaluator and generate train/test edge split
traintest_split = EvalSplit(
) # Bhevencious: EvalSplit() contains methods used to READ/SET a variety of properties/variables. Use the DOT & PARANTHESIS helpers to access parameters.
traintest_split.compute_splits(G,
nw_name='Zachary-Karate.edgelist',
train_frac=0.8)
nee = LPEvaluator(traintest_split)
# Create a Scoresheet to store the results
scoresheet = Scoresheet()
# Set the baselines
methods = [
'adamic_adar_index', 'common_neighbours', 'jaccard_coefficient', 'katz',
'preferential_attachment', 'resource_allocation_index', 'random_prediction'
]
# Evaluate baselines
for method in methods:
# Contact: [email protected] # Date: 18/12/2018 # This simple example is the one presented in the README.md file. from evalne.evaluation.evaluator import LPEvaluator from evalne.evaluation.split import EvalSplit from evalne.evaluation.score import Scoresheet from evalne.utils import preprocess as pp # Load and preprocess the network G = pp.load_graph('../evalne/tests/data/network.edgelist') G, _ = pp.prep_graph(G) # Create an evaluator and generate train/test edge split traintest_split = EvalSplit() traintest_split.compute_splits(G) nee = LPEvaluator(traintest_split) # Create a Scoresheet to store the results scoresheet = Scoresheet() # Set the baselines methods = ['random_prediction', 'common_neighbours', 'jaccard_coefficient'] # Evaluate baselines for method in methods: result = nee.evaluate_baseline(method=method) scoresheet.log_results(result) try: