def evolvegcn_embedding(dataset, learning_type='unsupervise'):
base_path = os.path.abspath(
os.path.join(os.getcwd(), '../data/' + dataset + '/CTGCN'))
origin_folder = os.path.join('..', '1.format')
origin_base_path = os.path.abspath(os.path.join(base_path, origin_folder))
embedding_folder = os.path.join('..', '2.embedding/EvolveGCNH')
node_file = os.path.join('..', 'nodes_set/nodes.csv')
duration = 15
max_time_num = len(os.listdir(origin_base_path))
node_path = os.path.abspath(os.path.join(base_path, node_file))
nodes_set = pd.read_csv(node_path, names=['node'])
data_loader = DataLoader(nodes_set['node'].tolist(), max_time_num)
t1 = time.time()
print('start EvolveGCN embedding!')
if learning_type == 'unsupervise':
walk_pair_folder = 'evolvegcn_walk_pairs'
node_freq_folder = 'evolvegcn_node_freq'
walk_pair_base_path = os.path.abspath(
os.path.join(base_path, walk_pair_folder))
node_freq_base_path = os.path.abspath(
os.path.join(base_path, node_freq_folder))
for idx in range(0, max_time_num, duration):
print('idx = ', idx)
adj_list = data_loader.get_date_adj_list(origin_base_path,
start_idx=idx,
duration=duration)
x_list, max_degree, _ = data_loader.get_degree_feature_list(
origin_base_path, start_idx=idx, duration=duration)
node_pair_list = data_loader.get_node_pair_list(
walk_pair_base_path, start_idx=idx, duration=duration)
neg_freq_list = data_loader.get_neg_freq_list(node_freq_base_path,
start_idx=idx,
duration=duration)
evolvegcn_model = EvolveGCN(input_dim=max_degree,
hidden_dim=128,
output_dim=128,
duration=duration,
egcn_type='EGCNH')
evolvegcn_loss = UnsupervisedLoss(neg_num=20,
Q=20,
node_pair_list=node_pair_list,
neg_freq_list=neg_freq_list)
evolvegcn = UnsupervisedEmbedding(
base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=evolvegcn_model,
loss=evolvegcn_loss,
max_time_num=max_time_num)
evolvegcn.learn_embedding(adj_list,
x_list,
epoch=5,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='evolvegcnh',
export=True)
break
elif learning_type == 'supervise':
label_file = os.path.join('..', 'nodes_set/trans_label.csv')
label_path = os.path.abspath(os.path.join(base_path, label_file))
df_label = pd.read_csv(label_path, sep='\t')
label_list = df_label['label'].values
for idx in range(0, max_time_num, duration):
print('idx = ', idx)
adj_list = data_loader.get_date_adj_list(origin_base_path,
start_idx=idx,
duration=duration)
x_list, max_degree, _ = data_loader.get_degree_feature_list(
origin_base_path, start_idx=idx, duration=duration)
evolvegcn_model = EvolveGCN(input_dim=max_degree,
hidden_dim=128,
output_dim=128,
duration=duration,
egcn_type='EGCNH')
evolvegcn_loss = SupervisedLoss()
evolvegcn_classifier = MLPClassifier(128,
64,
label_list.max() + 1,
layer_num=1,
duration=duration,
bias=True,
trans_version='L')
evolvegcn = SupervisedEmbedding(
base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=evolvegcn_model,
loss=evolvegcn_loss,
classifier=evolvegcn_classifier,
max_time_num=max_time_num)
evolvegcn.learn_embedding(adj_list,
x_list,
label_list,
epoch=50,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='evolvegcnh',
classifier_file='evolvegcnh_cls',
export=True)
else:
raise AttributeError('Unsupported learning type!')
t2 = time.time()
print('finish EvolveGCN embedding! cost time: ', t2 - t1, ' seconds!')
return
def ctgcn_connective_embedding(dataset, learning_type='unsupervise'):
base_path = os.path.abspath(
os.path.join(os.getcwd(), '../data/' + dataset + '/CTGCN'))
origin_folder = os.path.join('..', '1.format')
embedding_folder = os.path.join('..', '2.embedding/CTGCN_C')
core_folder = 'ctgcn_cores'
core_base_path = os.path.abspath(os.path.join(base_path, core_folder))
node_file = os.path.join('..', 'nodes_set/nodes.csv')
duration = 5
max_time_num = len(os.listdir(core_base_path))
node_path = os.path.abspath(os.path.join(base_path, node_file))
nodes_set = pd.read_csv(node_path, names=['node'])
node_num = nodes_set.shape[0]
data_loader = DataLoader(nodes_set['node'].tolist(), max_time_num)
print('max time num: ', max_time_num)
t1 = time.time()
print('start CTGCN_C embedding on ' + dataset)
if learning_type == 'unsupervise':
walk_pair_folder = 'ctgcn_walk_pairs'
node_freq_folder = 'ctgcn_node_freq'
walk_pair_base_path = os.path.abspath(
os.path.join(base_path, walk_pair_folder))
node_freq_base_path = os.path.abspath(
os.path.join(base_path, node_freq_folder))
for idx in range(0, max_time_num, duration):
print('idx = ', idx)
time_num = min(duration, max_time_num - idx)
adj_list = data_loader.get_core_adj_list(core_base_path,
start_idx=idx,
duration=time_num)
x_list = data_loader.get_feature_list(None,
start_idx=idx,
duration=time_num)
node_pair_list = data_loader.get_node_pair_list(
walk_pair_base_path, start_idx=idx, duration=time_num)
neg_freq_list = data_loader.get_neg_freq_list(node_freq_base_path,
start_idx=idx,
duration=time_num)
ctgcn_model = CTGCN(input_dim=node_num,
hidden_dim=500,
output_dim=128,
trans_num=1,
diffusion_num=2,
duration=time_num,
bias=True,
rnn_type='GRU',
version='C',
trans_version='L')
ctgcn_loss = UnsupervisedLoss(neg_num=150,
Q=10,
node_pair_list=node_pair_list,
neg_freq_list=neg_freq_list)
ctgcn = UnsupervisedEmbedding(base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=ctgcn_model,
loss=ctgcn_loss,
max_time_num=max_time_num)
ctgcn.learn_embedding(adj_list,
x_list,
single_output=False,
epoch=5,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='ctgcn_c',
embedding_type='connection',
export=True)
break
elif learning_type == 'supervise':
label_file = os.path.join('..', 'nodes_set/trans_label.csv')
label_path = os.path.abspath(os.path.join(base_path, label_file))
df_label = pd.read_csv(label_path, sep='\t')
label_list = df_label['label'].values
for idx in range(0, max_time_num, duration):
print('idx = ', idx)
time_num = min(duration, max_time_num - idx)
adj_list = data_loader.get_core_adj_list(core_base_path,
start_idx=idx,
duration=time_num)
x_list = data_loader.get_feature_list(None,
start_idx=idx,
duration=time_num)
ctgcn_model = CTGCN(input_dim=node_num,
hidden_dim=500,
output_dim=128,
trans_num=1,
diffusion_num=2,
duration=time_num,
bias=True,
rnn_type='GRU',
version='C',
trans_version='L')
ctgcn_loss = SupervisedLoss()
ctgcn_classifier = MLPClassifier(128,
64,
label_list.max() + 1,
layer_num=1,
duration=time_num,
bias=True,
trans_version='L')
ctgcn = SupervisedEmbedding(base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=ctgcn_model,
loss=ctgcn_loss,
classifier=ctgcn_classifier,
max_time_num=max_time_num)
ctgcn.learn_embedding(adj_list,
x_list,
label_list,
single_output=False,
epoch=50,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='ctgcn_c',
classifier_file='ctgcn_c_cls',
embedding_type='connection',
export=True)
else:
raise AttributeError('Unsupported learning type!')
t2 = time.time()
print('finish CTGCN_C embedding! cost time: ', t2 - t1, ' seconds!')
return
def gcn_embedding(dataset, learning_type='unsupervise'):
base_path = os.path.abspath(
os.path.join(os.getcwd(), '../data/' + dataset + '/CTGCN'))
origin_folder = os.path.join('..', '1.format')
origin_base_path = os.path.abspath(os.path.join(base_path, origin_folder))
embedding_folder = os.path.join('..', '2.embedding/GCN')
node_file = os.path.join('..', 'nodes_set/nodes.csv')
duration = 1
max_time_num = len(os.listdir(origin_base_path))
node_path = os.path.abspath(os.path.join(base_path, node_file))
nodes_set = pd.read_csv(node_path, names=['node'])
node_num = nodes_set.shape[0]
data_loader = DataLoader(nodes_set['node'].tolist(), max_time_num)
t1 = time.time()
print('start GCN embedding!')
if learning_type == 'unsupervise':
walk_pair_folder = 'gcn_walk_pairs'
node_freq_folder = 'gcn_node_freq'
walk_pair_base_path = os.path.abspath(
os.path.join(base_path, walk_pair_folder))
node_freq_base_path = os.path.abspath(
os.path.join(base_path, node_freq_folder))
# time_list = []
for idx in range(max_time_num):
print('idx = ', idx)
adj_list = data_loader.get_date_adj_list(origin_base_path,
start_idx=idx,
duration=duration)
x_list = data_loader.get_feature_list(None,
start_idx=idx,
duration=duration)
node_pair_list = data_loader.get_node_pair_list(
walk_pair_base_path, start_idx=idx, duration=duration)
neg_freq_list = data_loader.get_neg_freq_list(node_freq_base_path,
start_idx=idx,
duration=duration)
gcn_model = GCN(input_dim=node_num,
hidden_dim=500,
output_dim=128,
dropout=0.5,
bias=True)
gcn_loss = UnsupervisedLoss(neg_num=20,
Q=20,
node_pair_list=node_pair_list,
neg_freq_list=neg_freq_list)
gcn = UnsupervisedEmbedding(base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=gcn_model,
loss=gcn_loss,
max_time_num=max_time_num)
# ta = time.time()
gcn.learn_embedding(adj_list,
x_list,
epoch=5,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='gcn',
export=True)
if idx == 1:
break
# tb = time.time()
# time_list.append(tb - ta)
# df_output = pd.DataFrame({'time': time_list})
# df_output.to_csv(dataset + '_gcn_time.csv', sep=',', index=False)
elif learning_type == 'supervise':
label_file = os.path.join('..', 'nodes_set/trans_label.csv')
label_path = os.path.abspath(os.path.join(base_path, label_file))
df_label = pd.read_csv(label_path, sep='\t')
label_list = df_label['label'].values
for idx in range(max_time_num):
print('idx = ', idx)
adj_list = data_loader.get_date_adj_list(origin_base_path,
start_idx=idx,
duration=duration)
x_list = data_loader.get_feature_list(None,
start_idx=idx,
duration=duration)
gcn_model = GCN(input_dim=node_num,
hidden_dim=500,
output_dim=128,
dropout=0.5,
bias=True)
gcn_loss = SupervisedLoss()
gcn_classifier = MLPClassifier(128,
64,
label_list.max() + 1,
layer_num=1,
duration=duration,
bias=True,
trans_version='L')
gcn = SupervisedEmbedding(base_path=base_path,
origin_folder=origin_folder,
embedding_folder=embedding_folder,
node_list=nodes_set['node'].tolist(),
model=gcn_model,
loss=gcn_loss,
classifier=gcn_classifier,
max_time_num=max_time_num)
gcn.learn_embedding(adj_list,
x_list,
label_list,
epoch=50,
batch_size=4096 * 8,
lr=0.001,
start_idx=idx,
weight_decay=5e-4,
model_file='gcn',
classifier_file='gcn_cls',
export=True)
else:
raise AttributeError('Unsupported learning type!')
t2 = time.time()
print('finish GCN embedding! cost time: ', t2 - t1, ' seconds!')
return