def prepare_test_adj(self, input_graph, ground_truth_adj):
coords, values, shape = sparse_to_tuple(input_graph)
ground_truth_adj = (ground_truth_adj[:input_graph.shape[0], :input_graph.shape[1]]).todense()
for coord in coords:
ground_truth_adj[coord[0], coord[1]] = 0.
ground_truth_adj[coord[1], coord[0]] = 0.
return sps.triu(sps.csr_matrix(ground_truth_adj, dtype=float))
def construct_dataset(self, graph, window_size, negative_sample):
start_graph = max(0, graph - window_size + 1)
max_id = 0
for i in range(start_graph, graph + 1):
adj = self.graph_loader.read_adjacency(i, max_id)
max_id = adj.shape[0] - 1
train_adj_sps = []
total_train_edges = np.zeros((max_id + 1, max_id + 1))
for i in range(start_graph, graph + 1):
adj = self.graph_loader.read_adjacency(i, max_id)
tmp_train_adj_dense = adj.todense()
tmp_train_adj_dense = np.where(tmp_train_adj_dense > 0.2,
tmp_train_adj_dense, 0)
tmp_train_adj_sparse = sps.csr_matrix(tmp_train_adj_dense)
coords, values, shape = sparse_to_tuple(tmp_train_adj_sparse)
for coord in coords:
total_train_edges[coord[0], coord[1]] = 1
train_adj_sps.append(tmp_train_adj_sparse)
# Construct a full matrix with ones to generate negative sample tuples
train_ns = np.ones_like(
total_train_edges) - total_train_edges - sps.identity(
total_train_edges.shape[0])
ns_coord, ns_values, ns_shape = sparse_to_tuple(train_ns)
train_adj_norm = []
features = []
train_adj_labels = []
train_adj_inds = []
features_tuples = sparse_to_tuple(
sps.identity(adj.shape[0], dtype=np.float32, format='coo'))
for i, adj in enumerate(train_adj_sps):
adj_norm_coord, adj_norm_values, adj_norm_shape = sparse_to_tuple(
self.normalize(adj))
train_adj_norm.append(
tf.SparseTensor(
indices=adj_norm_coord,
values=np.array(adj_norm_values, dtype='float32'),
dense_shape=[adj_norm_shape[0], adj_norm_shape[1]]))
features.append(
tf.SparseTensor(
indices=features_tuples[0],
values=features_tuples[1],
dense_shape=[features_tuples[2][0],
features_tuples[2][1]]))
tmp_train_adj_dense = adj.todense()
train_coord, train_values, train_shape = sparse_to_tuple(adj)
tmp_train_adj_ind = np.zeros_like(tmp_train_adj_dense)
sequence = [i for i in range(len(ns_coord))]
random_coords = set(
random.sample(sequence, negative_sample * len(train_coord)))
for coord in train_coord:
tmp_train_adj_ind[coord[0], coord[1]] = 1
for coord in random_coords:
tmp_train_adj_ind[ns_coord[coord][0], ns_coord[coord][1]] = 1
nnz_ind = np.nonzero(tmp_train_adj_ind)
tmp_train_label_val = tmp_train_adj_dense[nnz_ind]
train_adj_label_tensor = tf.convert_to_tensor(tmp_train_label_val,
dtype=tf.float32)
train_adj_labels.append(train_adj_label_tensor)
ind_list = []
for i in range(len(nnz_ind[0])):
ind_list.append([nnz_ind[0][i], nnz_ind[1][i]])
train_adj_inds.append(
tf.convert_to_tensor(ind_list, dtype=tf.int32))
test_adj_dense = self.prepare_test_adj(
sps.csr_matrix(total_train_edges),
self.graph_loader.read_adjacency(graph + 1, max_id)).todense()
test_adj_high = np.where(test_adj_dense > 0.2, test_adj_dense, 0)
test_adj_ind = np.where(test_adj_high > 0., 1, 0)
nnz_ind = np.nonzero(test_adj_ind)
ind_list = []
for i in range(len(nnz_ind[0])):
ind_list.append([nnz_ind[0][i], nnz_ind[1][i]])
test_adj = tf.convert_to_tensor(test_adj_high[nnz_ind],
dtype=tf.float32)
test_adj_ind = tf.convert_to_tensor(ind_list, dtype=tf.int32)
return train_adj_norm, train_adj_labels, train_adj_inds, features, test_adj, test_adj_ind
def construct_dataset(self, graph, window_size, negative_sample):
start_graph = max(0, graph - window_size)
max_id = 0
for i in range(start_graph, graph + 1):
adj = self.graph_loader.read_adjacency(i, max_id)
max_id = adj.shape[0] - 1
train_adj = []
total_train_edges = np.zeros((max_id + 1, max_id + 1))
for i in range(start_graph, graph + 1):
adj = self.graph_loader.read_adjacency(i, max_id)
train_adj_dense = adj.todense()
train_adj_dense = np.where(train_adj_dense > 0.2, train_adj_dense,
0)
train_adj_sparse = sps.csr_matrix(train_adj_dense)
coords, values, shape = sparse_to_tuple(train_adj_sparse)
for coord in coords:
total_train_edges[coord[0], coord[1]] = 1
train_adj.append(train_adj_sparse)
train_adj_norm = []
features = []
for i, adj in enumerate(train_adj):
train_adj_norm.append(
sparse_mx_to_torch_sparse_tensor(
self.normalize(adj)).to(device=self.device))
features.append(
torch.tensor(
sps.identity(adj.shape[0], dtype=np.float32,
format='coo').todense(),
dtype=torch.float32).to_sparse().to(device=self.device))
# Generate the train_adj_label with negative sampling
if i == len(train_adj) - 1:
# Construct a full matrix with ones to generate negative sample tuples
train_ns = np.ones_like(
adj.todense()) - total_train_edges - sps.identity(
total_train_edges.shape[0])
ns_coord, ns_values, ns_shape = sparse_to_tuple(train_ns)
train_coord, train_values, train_shape = sparse_to_tuple(adj)
train_label_ind = np.zeros_like(adj.todense())
for coord in train_coord:
train_label_ind[coord[0], coord[1]] = 1
sequence = [i for i in range(len(ns_coord))]
random_coords = set(
random.sample(sequence,
negative_sample * len(train_coord)))
for coord in random_coords:
train_label_ind[ns_coord[coord][0], ns_coord[coord][1]] = 1
train_adj_dense = adj.todense()
nnz_ind = np.nonzero(train_label_ind)
train_label_val = train_adj_dense[nnz_ind][0]
train_adj_label = torch.reshape(
torch.tensor(train_label_val,
dtype=torch.float32).to(device=self.device),
(-1, ))
train_adj_ind = [
torch.tensor(nnz_ind[0],
requires_grad=False).to(device=self.device),
torch.tensor(nnz_ind[1],
requires_grad=False).to(device=self.device)
]
test_adj_dense = self.prepare_test_adj(
total_train_edges,
self.graph_loader.read_adjacency(graph + 1, max_id)).todense()
test_adj_high = np.where(test_adj_dense > 0.2, test_adj_dense, 0)
test_adj_ind = np.where(test_adj_high > 0., 1, 0)
nnz_ind = np.nonzero(test_adj_ind)
test_adj = torch.tensor(test_adj_high[nnz_ind],
dtype=torch.float32,
requires_grad=False).to(device=self.device)
test_adj_ind = [
torch.tensor(nnz_ind[0],
requires_grad=False).to(device=self.device),
torch.tensor(nnz_ind[1],
requires_grad=False).to(device=self.device)
]
return train_adj_norm, train_adj_label, train_adj_ind, features, test_adj, test_adj_ind