-
Notifications
You must be signed in to change notification settings - Fork 0
/
DC2B.py
285 lines (238 loc) · 10.1 KB
/
DC2B.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
274
275
276
277
278
279
280
281
282
283
284
import argparse
import os
import numpy as np
from utils import logging
from collections import defaultdict
import time
from sklearn.metrics.pairwise import cosine_similarity
import pdb
def parameter_setting():
parser = argparse.ArgumentParser()
parser.add_argument('--is_log', default=False)
parser.add_argument('--random_state', default=0)
parser.add_argument('--verbose', default=2)
parser.add_argument('--lam_da', default=1)
parser.add_argument('--sigma', default=1.)
parser.add_argument('--num_recommendation', default=10)
parser.add_argument('--hidden_dim', default=10)
parser.add_argument('--user_dim', default=4831)
parser.add_argument('--movie_dim', default=3496)
parser.add_argument('--num_bandit_iter', default=10)
parser.add_argument('--dpp_theta', default=0.5)
parser.add_argument('--dpp_w_size', default=10)
return parser.parse_args()
def f_train_ratings():
file_name = "ml_1m_user_new/ml-1m_user_0.8_train.txt"
user_history = defaultdict(lambda: defaultdict(int))
with open(file_name, 'r') as inf:
for line in inf:
data = line.split("\t")
u, i, r = data[0], data[1], data[2]
user_history[int(u)][int(i)] = float(r)
return user_history
def f_user_ratings():
file_name = "ml_1m_user_new/ml-1m_user_0.8_test.txt"
user_history = defaultdict(lambda: defaultdict(int))
with open(file_name, 'r') as inf:
for line in inf:
data = line.split("\t")
u, i, r = data[0], data[1], data[2]
user_history[int(u)][int(i)] = float(r)
return user_history
def rho(x):
return np.where(x > 0, 1. / (1. + np.exp(-x)), np.exp(x) / (np.exp(x) + 1.))
def lam_da(x):
return 1./(4*x)*np.tanh(x/2.)
def diversity(s_inx, X):
"""
s_inx: the selected item set, a numpy vector
X: the item feature matrix, shape (d, m)
The similarities between items are measured using cosine similarity.
"""
S = cosine_similarity(X[:, s_inx].T)
ii, jj = np.triu_indices(len(s_inx), k=1)
vec = S[ii, jj]
div = 1 - np.mean(vec)
return div
def bayes_greedy_map(scores, movie_embs, K, theta):
"""
all_movie_embs: m * d
"""
C = defaultdict(list)
# alpha = theta / (1 - theta)
alpha = theta
# vec = np.sqrt(rho(scores))
vec = np.exp(alpha * scores)
D = vec * vec
D = D * np.sum(movie_embs * movie_embs, axis=1)
j = np.argmax(D * rho(scores))
Y = [j] # the selected set
m = scores.shape[0]
Z = set(range(m)) # the remained items
for k in range(1, K):
Z = Z - set(Y)
for i in Z:
# define similarity matrix
Sji = (np.sum(movie_embs[j] * movie_embs[i]) + 1) / 2
# define L kernel
Lji = vec[j] * vec[i] * Sji
# Lji = Sji
if len(C[i]) == 0 or len(C[j]) == 0:
ei = Lji / (D[j] ** 0.5)
else:
ei = (Lji - np.sum(np.array(C[i]) * np.array(C[j]))) / (D[j] ** 0.5)
C[i].append(ei)
D[i] = D[i] - ei ** 2
ii = np.array(list(Z))
# greedy search for the next item
jj = np.argmax(D[ii] * rho(scores[ii]))
j = ii[jj]
Y.append(j)
# pdb.set_trace()
return Y
def mf_recommendation(user_emb, movie_embs, can_items, size=5):
inx = np.array(list(can_items))
scores = np.dot(user_emb, movie_embs[:, inx])
ii = np.argsort(scores)[::-1][:size]
return list(inx[ii])
def pre_diversity(s_inx, item_sim):
num = len(s_inx)
s = []
for i in range(num):
for j in range(i+1, num):
s.append(1 - item_sim[s_inx[i], s_inx[j]])
return sum(s) / len(s)
def bayesian_dpp(embeddings, test_items, args,
num=10, lamb_da=0.1,
train_items=None, user_emb=None, cate_sim=None):
"""
user_emb: user embedding, shape (d, 1)
movie_embs: movie embeddings, shape (d, m)
"""
hidden_dim, num_movies = embeddings.shape
matrix_s = lamb_da * np.identity(hidden_dim, dtype=np.float32)
vector_m = np.ones(shape=hidden_dim, dtype=np.float32)
prec = []
recall = []
div = []
cate_div = []
rec_items = []
all_items = set(range(num_movies))
if train_items is None:
can_items = all_items
else:
can_items = all_items - set(train_items)
# feature normalization
nor_embs = embeddings.T.copy()
for i in range(num_movies):
nor_embs[i, :] = nor_embs[i, :] / np.linalg.norm(nor_embs[i, :])
for t in range(num):
theta_hat = vector_m
p_hat = np.dot(theta_hat, embeddings)
can_items = can_items - set(rec_items)
kk = np.array(list(can_items))
if t == 0 and user_emb is not None:
s_inx = mf_recommendation(user_emb, embeddings, can_items, size=5)
else:
# get recommendation set s then delete it from the candidate items
s_tmp = bayes_greedy_map(p_hat[kk], nor_embs[kk, :], args.num_recommendation, args.dpp_theta)
s_inx = [kk[i] for i in s_tmp]
rec_items.extend(s_inx)
x = embeddings[:, np.array(s_inx)]
m = vector_m.reshape(hidden_dim, 1)
xi_tmp = np.dot(x.T, matrix_s+np.dot(m, m.T)) # shape=(m_s, d)
xi = np.sqrt(np.sum(xi_tmp.T*x, axis=0)) # shape=(m_s)
inv_matrix_s = np.linalg.inv(matrix_s)
lam_xi = np.tile(lam_da(xi), (hidden_dim, 1))
s_i = np.sum((lam_xi*x).T[:, None, :]*x.T[:, :, None], axis=0)
# here we get updated covariance matrix S
inv_matrix_s_post = inv_matrix_s + 2*s_i
matrix_s = np.linalg.inv(inv_matrix_s_post)
# here we get our reward Y
reward = np.array([1.0 if i in test_items else 0.0 for i in s_inx])
m_i = np.sum(np.tile((reward+3./2), (hidden_dim, 1))*x, axis=1)
# here we get updated mean vector m
vector_m_post = np.dot(inv_matrix_s, vector_m)+m_i
vector_m = np.dot(matrix_s, vector_m_post)
# precision for each iteration
inter_set = set(s_inx).intersection(set(list(test_items.keys())))
prec_curr = float(len(inter_set)) / float(args.num_recommendation)
prec.append(prec_curr)
# compute recall
s_test = set(list(test_items.keys()))
recall_curr = float(len(inter_set)) / float(len(s_test))
recall.append(recall_curr)
# calculate the diversity
div_curr = diversity(s_inx, embeddings)
div.append(div_curr)
if cate_sim is None:
cate_div_cur = 0.0
else:
cate_div_cur = pre_diversity(s_inx, cate_sim)
cate_div.append(cate_div_cur)
return np.array(prec), np.array(recall), np.array(div), np.array(cate_div)
if __name__ == '__main__':
args = parameter_setting()
# logging
if args.is_log:
file_name = os.path.basename(__file__)
output_path = logging(file_name, verbose=2)
print(args)
test_user_ratings = f_user_ratings()
train_user_ratings = f_train_ratings()
movie_cate_sim = np.load("ml_1m_0.8/ml-1m_tmp_0.8_10_item_sim.npy")
user_embs = np.load("ml_1m_user_new/bpr_ml-1m_user_0.8_dim10_user_embs.npy").T
movie_embs = np.load("ml_1m_user_new/bpr_ml-1m_user_0.8_dim10_item_embs.npy").T
args.user_dim = user_embs.shape[1]
args.movie_dim = movie_embs.shape[1]
sim_mat = np.dot(movie_embs.T, movie_embs)
for theta in [3]:
print(theta)
l = 1.
test_precision = np.zeros(args.num_bandit_iter)
test_recall = np.zeros(args.num_bandit_iter)
test_diversity = np.zeros(args.num_bandit_iter)
test_cate_diversity = np.zeros(args.num_bandit_iter)
test_precision_low = np.zeros(args.num_bandit_iter)
test_recall_low = np.zeros(args.num_bandit_iter)
test_diversity_low = np.zeros(args.num_bandit_iter)
test_cate_diversity_low = np.zeros(args.num_bandit_iter)
args.dpp_theta = theta
t1 = time.clock()
length = 0.
length_low = 0.
for user in test_user_ratings.keys():
if len(test_user_ratings[user]) >= 20:
print(user)
mv_embs = movie_embs.copy()
prec, rec, div, cate_div = bayesian_dpp(mv_embs, test_user_ratings[user], args,
num=args.num_bandit_iter, lamb_da=l,
train_items=None,
user_emb=None,
cate_sim=movie_cate_sim)
print(prec)
test_precision += prec
test_recall += rec
test_diversity += div
test_cate_diversity += cate_div
length += 1
else:
mv_embs = movie_embs.copy()
prec, rec, div, cate_div = bayesian_dpp(mv_embs, test_user_ratings[user], args,
num=args.num_bandit_iter, lamb_da=l,
train_items=None,
user_emb=None,
cate_sim=movie_cate_sim)
test_precision_low += prec
test_recall_low += rec
test_diversity_low += div
test_cate_diversity_low += cate_div
length_low += 1
print("theta:{0}, ".format(theta))
print("test_precision:{0}".format(test_precision / length))
print("test_recall:{0}".format(test_recall / length))
print("test_diversity:{0}".format(test_diversity / length))
print("test_cate_diversity:{0}".format(test_cate_diversity / length))
print("test_precision_low:{0}".format(test_precision_low / length_low))
print("test_precision_all:{0}".format((test_precision+test_precision_low)/(length+length_low)))
print("time used:%s" % (time.clock() - t1))