def hybrid_logistic(dataset, k, embeddings, model):
print("Hybrid model: Using features..")
train_X, train_y, test_X, test_y = read_train_test_data(dataset, k)
fea = FeaExtra(k=k, dataset=dataset)
train_X1 = []
test_X1 = []
for i, j in train_X:
features_i_j = np.array(fea.get_features(i, j))
train_X1.append(
np.concatenate([embeddings[i], embeddings[j], features_i_j]))
for i, j in test_X:
features_i_j = np.array(fea.get_features(i, j))
test_X1.append(
np.concatenate([embeddings[i], embeddings[j], features_i_j]))
#print(train_X1[1])
logistic_function = linear_model.LogisticRegression()
logistic_function.fit(train_X1, train_y)
pred = logistic_function.predict(test_X1)
pred_p = logistic_function.predict_proba(test_X1)
pos_ratio = np.sum(test_y) / test_y.shape[0]
accuracy = metrics.accuracy_score(test_y, pred)
f1_score0 = metrics.f1_score(test_y, pred)
f1_score1 = metrics.f1_score(test_y, pred, average='macro')
f1_score2 = metrics.f1_score(test_y, pred, average='micro')
auc_score = metrics.roc_auc_score(test_y, pred_p[:, 1])
print("pos_ratio:", pos_ratio)
print('accuracy:', accuracy)
print("f1_score:", f1_score0)
print("macro f1_score:", f1_score1)
print("micro f1_score:", f1_score2)
print("auc score:", auc_score)
return pos_ratio, accuracy, f1_score0, f1_score1, f1_score2, auc_score
def logistic_embedding8(k=1, dataset='epinions'):
"""use feature to train logistic function
Returns:
pos_ratio, accuracy, f1_score0, f1_score1, f1_score2, auc_score
"""
print(dataset, k, 'fea')
train_X, train_y, test_X, test_y = read_train_test_data(dataset, k)
fea = FeaExtra(k=k, dataset=dataset)
train_X1 = []
test_X1 = []
for i, j in train_X:
train_X1.append(fea.get_features(i, j))
for i, j in test_X:
test_X1.append(fea.get_features(i, j))
logistic = linear_model.LogisticRegression()
logistic.fit(train_X1, train_y)
pred = logistic.predict(test_X1)
pred_p = logistic.predict_proba(test_X1)
pos_ratio = np.sum(test_y) / test_y.shape[0]
accuracy = metrics.accuracy_score(test_y, pred)
f1_score0 = metrics.f1_score(test_y, pred)
f1_score1 = metrics.f1_score(test_y, pred, average='macro')
f1_score2 = metrics.f1_score(test_y, pred, average='micro')
auc_score = metrics.roc_auc_score(test_y, pred_p[:, 1])
print("pos_ratio:", pos_ratio)
print('accuracy:', accuracy)
print("f1_score:", f1_score0)
print("macro f1_score:", f1_score1)
print("micro f1_score:", f1_score2)
print("auc score:", auc_score)
return pos_ratio, accuracy, f1_score0, f1_score1, f1_score2, auc_score
def run(dataset='bitcoin_alpha', k=2):
num_nodes = DATASET_NUM_DIC[dataset] + 3
filename = './experiment-data/{}-train-{}.edgelist'.format(dataset, k)
adj_lists1, adj_lists1_1, adj_lists1_2, adj_lists2, adj_lists2_1, adj_lists2_2, adj_lists3 = load_data2(
filename, add_public_foe=False)
print(k, dataset, 'data load!')
features = nn.Embedding(num_nodes, NODE_FEAT_SIZE)
features.weight.requires_grad = True
features.to(DEVICES)
adj_lists = [
adj_lists1, adj_lists1_1, adj_lists1_2, adj_lists2, adj_lists2_1,
adj_lists2_2
]
#######
fea_model = FeaExtra(dataset=dataset, k=k)
adj_additions1 = [defaultdict(set) for _ in range(16)]
adj_additions2 = [defaultdict(set) for _ in range(16)]
adj_additions0 = [defaultdict(set) for _ in range(16)]
a, b = 0, 0
for i in adj_lists3:
for j in adj_lists3[i]:
v_list = fea_model.feature_part2(i, j)
for index, v in enumerate(v_list):
if v > 0:
adj_additions0[index][i].add(j)
for i in adj_lists1_1:
for j in adj_lists1_1[i]:
v_list = fea_model.feature_part2(i, j)
for index, v in enumerate(v_list):
if v > 0:
adj_additions1[index][i].add(j)
a += 1
for i in adj_lists2_1:
for j in adj_lists2_1[i]:
v_list = fea_model.feature_part2(i, j)
for index, v in enumerate(v_list):
if v > 0:
adj_additions2[index][i].add(j)
b += 1
assert a > 0, 'positive something wrong'
assert b > 0, 'negative something wrong'
adj_lists = adj_lists + adj_additions1 + adj_additions2
print(len(adj_lists), 'motifs')
def func(adj_list):
edges = []
for a in adj_list:
for b in adj_list[a]:
edges.append((a, b))
edges = np.array(edges)
adj = sp.csr_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(num_nodes, num_nodes))
return adj
adj_lists = list(map(func, adj_lists))
features_lists = [features for _ in range(len(adj_lists))]
aggs = [
AttentionAggregator(features, NODE_FEAT_SIZE, NODE_FEAT_SIZE,
num_nodes)
for features, adj in zip(features_lists, adj_lists)
]
enc1 = Encoder(features_lists, NODE_FEAT_SIZE, EMBEDDING_SIZE1, adj_lists,
aggs)
model = SiGAT(enc1)
model.to(DEVICES)
# print(model.train())
model.train()
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
list(model.parameters()) + list(enc1.parameters()) \
+ list(features.parameters())),
lr=LEARNING_RATE,
weight_decay=WEIGHT_DECAY
)
for epoch in range(EPOCHS + 2):
total_loss = []
if epoch % INTERVAL_PRINT == 0:
model.eval()
all_embedding = np.zeros((NUM_NODE, EMBEDDING_SIZE1))
for i in range(0, NUM_NODE, BATCH_SIZE):
begin_index = i
end_index = i + BATCH_SIZE if i + BATCH_SIZE < NUM_NODE else NUM_NODE
values = np.arange(begin_index, end_index)
embed = model.forward(values.tolist())
embed = embed.data.cpu().numpy()
all_embedding[begin_index:end_index] = embed
model.train()
time1 = time.time()
nodes_pku = np.random.permutation(NUM_NODE).tolist()
for batch in range(NUM_NODE // BATCH_SIZE):
optimizer.zero_grad()
b_index = batch * BATCH_SIZE
e_index = (batch + 1) * BATCH_SIZE
nodes = nodes_pku[b_index:e_index]
loss = model.criterion(nodes, adj_lists1, adj_lists2)
total_loss.append(loss.data.cpu().numpy())
loss.backward()
optimizer.step()
print(
f'epoch: {epoch}, loss: {np.sum(total_loss)}, time: {time.time()-time1}'
)
fpath = os.path.join(
OUTPUT_DIR, 'embedding-{}-{}-{}_{}.npy'.format(dataset, k,
str(EMBEDDING_SIZE1),
str(epoch - 1)))
np.save(fpath, all_embedding)
def run(dataset, k):
num_nodes = DATASET_NUM_DIC[dataset] + 3
# adj_lists1, adj_lists2, adj_lists3 = load_data(k, dataset)
filename = './experiment-data/{}-train-{}.edgelist'.format(dataset, k)
adj_lists1, adj_lists1_1, adj_lists1_2, adj_lists2, adj_lists2_1, adj_lists2_2, adj_lists3 = load_data2(
filename)
print(k, dataset, 'data load!')
features = nn.Embedding(num_nodes, NODE_FEAT_SIZE)
features.weight.requires_grad = True
features = features.to(DEVICES)
# 有向的四个边; 4个motifs
adj_lists = [adj_lists1_1, adj_lists1_2, adj_lists2_1, adj_lists2_2]
weight_dict = defaultdict(dict)
fea_model = FeaExtra(dataset=dataset, k=k)
# # u -> v
for i in adj_lists1_1:
for j in adj_lists1_1[i]:
v_list1 = fea_model.feature_part2(i, j)
mask = [1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1]
counts1 = np.dot(v_list1, mask) # 构造balance三角形，在后面计算loss时候使用;
weight_dict[i][j] = counts1
for i in adj_lists2_1:
for j in adj_lists2_1[i]:
v_list1 = fea_model.feature_part2(i, j)
mask = [0, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 1, 0, 0]
counts1 = np.dot(v_list1, mask)
weight_dict[i][j] = counts1
adj_lists = adj_lists
print(len(adj_lists), 'motifs')
def func(adj_list):
edges = []
for a in adj_list:
for b in adj_list[a]:
edges.append((a, b))
edges = np.array(edges)
adj = sp.csr_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(num_nodes, num_nodes))
return adj
if args.agg == 'mean':
aggregator = MeanAggregator
else:
aggregator = AttentionAggregator
adj_lists = list(map(func, adj_lists))
aggs = [
aggregator(features, NODE_FEAT_SIZE, NODE_FEAT_SIZE, num_nodes)
for adj in adj_lists
] # 构建motifs下节点的邻居聚合
enc1 = Encoder(features, NODE_FEAT_SIZE, EMBEDDING_SIZE1, adj_lists,
aggs) # 4个motifs+本身节点 + cls => embedding
enc1 = enc1.to(DEVICES)
aggs2 = [
aggregator(lambda n: enc1(n), EMBEDDING_SIZE1, EMBEDDING_SIZE1,
num_nodes) for _ in adj_lists
]
enc2 = Encoder(lambda n: enc1(n), EMBEDDING_SIZE1, EMBEDDING_SIZE1,
adj_lists, aggs2)
model = SDGNN(enc2)
model = model.to(DEVICES)
print(model.train())
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
list(model.parameters()) + list(enc1.parameters()) \
+ list(features.parameters())),
lr=LEARNING_RATE,
weight_decay=WEIGHT_DECAY
)
for epoch in range(EPOCHS + 2):
total_loss = []
if epoch % INTERVAL_PRINT == 1:
model.eval()
all_embedding = np.zeros((NUM_NODE, EMBEDDING_SIZE1))
for i in range(0, NUM_NODE, BATCH_SIZE):
begin_index = i
end_index = i + BATCH_SIZE if i + BATCH_SIZE < NUM_NODE else NUM_NODE
values = np.arange(begin_index, end_index)
embed = model.forward(values.tolist())
embed = embed.data.cpu().numpy()
all_embedding[begin_index:end_index] = embed
fpath = os.path.join(
OUTPUT_DIR,
'embedding-{}-{}-{}.npy'.format(dataset, k, str(epoch)))
np.save(fpath, all_embedding)
pos_ratio, accuracy, f1_score0, f1_score1, f1_score2, auc_score = logistic_embedding(
k=k, dataset=dataset, epoch=epoch, dirname=OUTPUT_DIR)
model.train()
time1 = time.time()
nodes_pku = np.random.permutation(NUM_NODE).tolist()
for batch in range(NUM_NODE // BATCH_SIZE):
optimizer.zero_grad()
b_index = batch * BATCH_SIZE
e_index = (batch + 1) * BATCH_SIZE
nodes = nodes_pku[b_index:e_index]
loss = model.criterion(
nodes,
adj_lists1,
adj_lists2,
adj_lists1_1,
adj_lists2_1,
weight_dict # i,j节点的平衡三角形数量
)
total_loss.append(loss.data.cpu().numpy())
loss.backward()
optimizer.step()
print(
f'epoch: {epoch}, loss: {np.mean(total_loss)}, time: {time.time()-time1}'
)