forked from LuoUndergradXJTU/TwiBot-22
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ours_twi20.py
273 lines (253 loc) · 12.2 KB
/
ours_twi20.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
import os.path as osp
import numpy as np
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import Linear, HGTConv
import heapq
import random
from scipy.spatial.distance import pdist
from sklearn.metrics import roc_auc_score
from torch_geometric.data import HeteroData
import scipy.io as sio
from sklearn.metrics import recall_score
from sklearn.metrics import precision_score
from sklearn.metrics import accuracy_score
from sklearn.metrics import f1_score
data2 = sio.loadmat(f'twi20.mat')
data1 = HeteroData()
data1['user'].x = torch.FloatTensor(data2['X1'])
data1['tweet'].x = torch.FloatTensor(data2['X2'][:,:16])
data1.node_types=['user', 'tweet']
data1.edge_types=[('user', 'to', 'tweet'), ('tweet', 'to', 'user')]
for i in range(data2['edge_index'].shape[1]):
data2['edge_index'][1, i] -= data2['X1'].shape[0]
data1[('user', 'to', 'tweet')].edge_index = torch.LongTensor(data2['edge_index'])
arr = np.copy(data2['edge_index'])
arr[[0, 1], :] = arr[[1, 0], :]
data1[('tweet', 'to', 'user')].edge_index = torch.LongTensor(arr)
# data1[('user', 'to', 'director')].edge_index = torch.LongTensor(data[('user', 'to', 'director')]['edge_index'])
# data1[('director', 'to', 'user')].edge_index = torch.LongTensor(data[('director', 'to', 'user')]['edge_index'])
node_types=['user', 'tweet']
gnd = data2['label'][0, :]
gnd1 = gnd[0:data2['X1'].shape[0]]
gnd2 = gnd[data2['X1'].shape[0]:]
T=dict()
T['user']=torch.zeros((data1['user'].x.shape[0],2))
T['tweet']=torch.zeros((data1['tweet'].x.shape[0],2))
for i in range(data1['user'].x.shape[0]):
T['user'][i,0]=1
for i in range(data1['tweet'].x.shape[0]):
T['tweet'][i,1]=1
X = dict()
for node_type in data1.node_types:
X[node_type] = data1[node_type].x
A = dict()
flag_user = []
for i in range(data1.x_dict['user'].shape[1]):
flag_user.append(random.randint(1, 4))
flag_user = np.array(flag_user)
flag_tweet = []
for i in range(data1.x_dict['tweet'].shape[1]):
flag_tweet.append(random.randint(1, 2))
flag_tweet = np.array(flag_tweet)
def multiview(x, num1, num2):
X_hat_views = dict()
X_hat_views['user'] = []
X_hat_views['tweet'] = []
index_user = []
index_tweet = []
n = 0
m = 0
for i in range(num1):
while n < len(flag_user):
if flag_user[n] == i + 1:
index_user.append(n)
n += 1
X_hat_views['user'].append(x['user'][:, index_user])
index_user = []
n=0
for i in range(num2):
while m < len(flag_tweet):
if flag_tweet[m] == i + 1:
index_tweet.append(m)
m += 1
X_hat_views['tweet'].append(x['tweet'][:, index_tweet])
index_tweet = []
m=0
return X_hat_views
X_views=multiview(X, 4, 2)
y_dict1 = dict.copy(data1.x_dict)
y_dict1['user'] = X_views['user'][0]
y_dict1['tweet'] = X_views['tweet'][0]
y_dict2 = dict.copy(data1.x_dict)
y_dict2['user'] = X_views['user'][1]
y_dict2['tweet'] = X_views['tweet'][0]
y_dict3 = dict.copy(data1.x_dict)
y_dict3['user'] = X_views['user'][2]
y_dict3['tweet'] = X_views['tweet'][0]
y_dict4 = dict.copy(data1.x_dict)
y_dict4['user'] = X_views['user'][3]
y_dict4['tweet'] = X_views['tweet'][0]
y_dict5 = dict.copy(data1.x_dict)
y_dict5['user'] = X_views['user'][0]
y_dict5['tweet'] = X_views['tweet'][1]
y_dict6 = dict.copy(data1.x_dict)
y_dict6['user'] = X_views['user'][1]
y_dict6['tweet'] = X_views['tweet'][1]
y_dict7 = dict.copy(data1.x_dict)
y_dict7['user'] = X_views['user'][2]
y_dict7['tweet'] = X_views['tweet'][1]
y_dict8 = dict.copy(data1.x_dict)
y_dict8['user'] = X_views['user'][3]
y_dict8['tweet'] = X_views['tweet'][1]
list1 = [y_dict1, y_dict2, y_dict3, y_dict4, y_dict5, y_dict6, y_dict7, y_dict8]
# A = torch.sparse.FloatTensor(i, v, torch.Size([4057, 14328])).to_dense()
for edge_type in data1.edge_types:
A[edge_type] = torch.sparse.FloatTensor(data1.edge_index_dict[edge_type],
torch.LongTensor(np.ones(data1.edge_index_dict[edge_type].shape[1])),
torch.Size([data1[edge_type[0]].x.shape[0],
data1[edge_type[2]].x.shape[0]])).to_dense()
class HGT(torch.nn.Module):
def __init__(self, hidden_channels, out_channels, num_heads, num_layers, num_view):
super().__init__()
self.weight = torch.nn.Parameter(torch.randn(num_view), requires_grad=True)
self.weight_node_type = torch.nn.Parameter(torch.randn(2), requires_grad=True)
# self.w_FC = torch.nn.Parameter(torch.randn(out_channels, data.x_dict['user'].shape[1]), requires_grad=True)
# self.b_FC = torch.nn.Parameter(torch.randn(data.x_dict['user'].shape), requires_grad=True)
self.lin_dict = torch.nn.ModuleDict()
for node_type in data1.node_types:
self.lin_dict[node_type] = torch.nn.ModuleList()
for i in range(8):
if i == 0:
self.lin_dict[node_type].append(Linear(y_dict1[node_type].shape[1], hidden_channels))
if i == 1:
self.lin_dict[node_type].append(Linear(y_dict2[node_type].shape[1], hidden_channels))
if i == 2:
self.lin_dict[node_type].append(Linear(y_dict3[node_type].shape[1], hidden_channels))
if i == 3:
self.lin_dict[node_type].append(Linear(y_dict4[node_type].shape[1], hidden_channels))
if i == 4:
self.lin_dict[node_type].append(Linear(y_dict5[node_type].shape[1], hidden_channels))
if i == 5:
self.lin_dict[node_type].append(Linear(y_dict6[node_type].shape[1], hidden_channels))
if i == 6:
self.lin_dict[node_type].append(Linear(y_dict7[node_type].shape[1], hidden_channels))
if i == 7:
self.lin_dict[node_type].append(Linear(y_dict8[node_type].shape[1], hidden_channels))
self.convs = torch.nn.ModuleList()
for _ in range(num_layers):
conv = HGTConv(hidden_channels, hidden_channels, data1.metadata(), num_heads, group='sum')
self.convs.append(conv)
self.out_dict = torch.nn.ModuleDict()
for node_type in data1.node_types:
self.out_dict[node_type] = Linear(hidden_channels, out_channels)
self.lin = torch.nn.ModuleDict()
for node_type in data1.node_types:
self.lin[node_type] = Linear(out_channels, data1.x_dict[node_type].shape[1])
self.Tlin = torch.nn.ModuleDict()
for node_type in data1.node_types:
self.Tlin[node_type] = Linear(data1.x_dict[node_type].shape[1], 2)
# self.Tlin= Linear(data.x_dict['user'].shape[1], 4)
def forward(self, x_dict, edge_index_dict):
list3 = dict()
for node_type, _ in x_dict[0].items():
list3[node_type] = []
for i in range(8):
for node_type, x in x_dict[i].items():
x_dict[i][node_type] = self.lin_dict[node_type][i](x).relu_()
for conv in self.convs:
x_dict[i] = conv(x_dict[i], edge_index_dict)
for node_type, _ in x_dict[0].items():
list3[node_type].append(self.out_dict[node_type](x_dict[i][node_type]))
weight_norm = F.softmax(self.weight, dim=0)
weight_type = F.softmax(self.weight_node_type, dim=0)
Z_dict = dict()
A_head = dict()
X_head = dict()
t_head = dict()
T_head = dict()
for node_type, _ in x_dict[0].items():
# Z_dict[node_type] = weight_norm[0] * list3[node_type][0] + weight_norm[1] * list3[node_type][1]
Z_dict[node_type] = weight_norm[0] * list3[node_type][0] + weight_norm[1] * list3[node_type][1] + \
weight_norm[2] * list3[node_type][2] + weight_norm[3] * list3[node_type][3] + \
weight_norm[4] * list3[node_type][4] + weight_norm[5] * list3[node_type][5] + \
weight_norm[6] * list3[node_type][6] + weight_norm[7] * list3[node_type][7]
X_head[node_type] = self.lin[node_type](Z_dict[node_type])
t_head[node_type] = self.Tlin[node_type](X_head[node_type])
T_head[node_type] = torch.zeros((t_head[node_type].shape[0], t_head[node_type].shape[1]))
for i in range(T_head[node_type].shape[0]):
T_head[node_type][i, 0] = weight_type[0] * t_head[node_type][i, 0]
T_head[node_type][i, 1] = weight_type[1] * t_head[node_type][i, 1]
T_head[node_type] = F.softmax(T_head[node_type], dim=0)
for edge_type in edge_index_dict.keys():
A_head[edge_type] = torch.sigmoid(torch.mm(Z_dict[edge_type[0]], Z_dict[edge_type[2]].T))
# Z_user = weight_norm[0] * list3[0] + weight_norm[1] * list3[1]
# Z_tweet = self.lin(x_dict[0]['tweet'])
# A_head = torch.sigmoid(torch.mm(Z_user, Z_tweet.T))
# X_head = torch.relu_(torch.mm(Z_user, self.w_FC) + self.b_FC)
# T_head = F.softmax(X_head , dim=1)
return A_head, X_head, T_head, weight_norm, weight_type
model = HGT(hidden_channels=128, out_channels=16, num_heads=2, num_layers=2, num_view=8)
device = torch.device('cpu')
data, model = data1.to(device), model.to(device)
'''
with torch.no_grad(): # Initialize lazy modules.
out = model(list1, data.edge_index_dict)
'''
optimizer = torch.optim.Adam(model.parameters(), lr=0.005, weight_decay=0.00001)
def train():
model.train()
optimizer.zero_grad()
input_list = copy.deepcopy(list1)
A_hat, X_hat, T_hat, wX, wT = model(input_list, data.edge_index_dict)
X_hat_view = multiview(X_hat, 4, 2)
loss = 0
loss += pow(torch.norm(A_hat[('user', 'to', 'tweet')] - A[('user', 'to', 'tweet')]), 2)
loss += pow(torch.norm(A_hat[('tweet', 'to', 'user')] - A[('tweet', 'to', 'user')]), 2)
# loss += F.cross_entropy(out[0][('user', 'to', 'director')], torch.LongTensor(gnd1))
# loss += F.cross_entropy(out[0][('director', 'to', 'user')], torch.LongTensor(np.zeros(data['director'].x.shape[0])))
# loss += torch.norm(out[1]['user'] - X['user'])/float(gnd1.shape[0])
# loss += torch.norm(out[1]['tweet'] - X['tweet'])/float(gnd2.shape[0])
# loss += torch.norm(out[1]['director'] - X['director'])
# for node_type in data1.node_types:
#loss += (torch.norm(out[2][node_type] - torch.FloatTensor(np.array(T[node_type]))))/data1[node_type].x.shape[0]
for node_type in node_types:
loss += pow(torch.norm(T_hat[node_type] - T[node_type]), 2)
for i in range(4):
loss += pow(torch.norm(X_hat_view['user'][i] - X_views['user'][i]), 2)
for i in range(2):
loss += pow(torch.norm(X_hat_view['tweet'][i] - X_views['tweet'][i]), 2)
loss=loss/float(gnd.shape[0])
loss.backward()
optimizer.step()
return float(loss)
@torch.no_grad()
def test():
model.eval()
input_list = copy.deepcopy(list1)
out = model(input_list, data.edge_index_dict)
ano_score = []
for i in range(data1['user'].x.shape[0]):
ano_score.append(0.48 * torch.norm(out[0][('user', 'to', 'tweet')][i] - A[('user', 'to', 'tweet')][i])
+ 0.48 * torch.norm(out[1]['user'][i] - X['user'][i])+0.04 * torch.norm(out[2]['user'][i] - torch.FloatTensor(np.array(T['user']))))
for i in range(data1['tweet'].x.shape[0]):
ano_score.append(0.48 * torch.norm(out[0][('tweet', 'to', 'user')][i] - A[('tweet', 'to', 'user')][i])
+ 0.48 * torch.norm(out[1]['tweet'][i] - X['tweet'][i])+0.04 * torch.norm(out[2]['tweet'][i] - torch.FloatTensor(np.array(T['tweet']))))
index_ano = list(map(ano_score.index, heapq.nlargest(int(np.sum(gnd)), ano_score)))
label_ano = np.zeros(gnd.shape[0])
for i in range(gnd.shape[0]):
if i in index_ano:
label_ano[i]=1
# acc = len(set(index_ano) & set(selected)) / 300.0
auc = roc_auc_score(gnd, np.array(ano_score) / max(ano_score))
Prec=precision_score(gnd, label_ano)
Rec=recall_score(gnd, label_ano)
F1=f1_score(gnd, label_ano)
acc=accuracy_score(gnd, label_ano)
return acc, Prec, Rec, F1, auc
for epoch in range(1, 101):
loss = train()
acc, Prec, Rec, F1, auc = test()
print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}, Acc: {acc:.4f}, Prec:{Prec:.4f}, Rec:{Rec:.4f}, F1:{F1:.4f}, AUC:{auc:.4f}')