/
pts.py
231 lines (199 loc) · 8.78 KB
/
pts.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
from threadpoolctl import threadpool_limits
import numpy as np
from tqdm import tqdm
import scipy.sparse
from collections import defaultdict
from .base import ValueFunction
from tqdm import tqdm
from numba import njit
import irec.recommendation.matrix_factorization as mf
@njit
def _softmax(x):
return np.exp(x - np.max(x)) / np.sum(np.exp(x - np.max(x)))
class PTS(ValueFunction):
"""Particle Thompson sampling.
It is a PMF formulation for the original TS based on a Bayesian inference around the items.
This method also applies particle filtering to guide the exploration of items over time [1]_.
References
----------
.. [1] Wang, Qing, et al. "Online interactive collaborative filtering using multi-armed
bandit with dependent arms." IEEE Transactions on Knowledge and Data Engineering 31.8 (2018): 1569-1580.
"""
def __init__(self, num_lat, num_particles, var, var_u, var_v, *args, **kwargs):
"""__init__.
Args:
args:
kwargs:
num_particles (int):
var:
var_u:
var_v:
"""
super().__init__(*args, **kwargs)
self.num_particles = num_particles
self.var = var
self.var_u = var_u
self.var_v = var_v
self.num_lat = num_lat
def reset(self, observation):
"""reset.
Args:
observation:
"""
train_dataset = observation
super().reset(train_dataset)
self.train_dataset = train_dataset
self.train_consumption_matrix = scipy.sparse.csr_matrix(
(
self.train_dataset.data[:, 2],
(self.train_dataset.data[:, 0], self.train_dataset.data[:, 1]),
),
(self.train_dataset.num_total_users, self.train_dataset.num_total_items),
)
self.num_total_items = self.train_dataset.num_total_items
self.num_total_users = self.train_dataset.num_total_users
self.particles_us = np.random.normal(
size=(self.num_particles, self.num_total_users, self.num_lat)
)
self.particles_vs = np.random.normal(
size=(self.num_particles, self.num_total_items, self.num_lat)
)
self.particles_var_us = np.ones(shape=(self.num_particles)) * self.var_u
self.particles_var_vs = np.ones(shape=(self.num_particles)) * self.var_v
self.particles_ids = np.arange(self.num_particles)
self.items_consumed_users = defaultdict(list)
self.items_consumed_users_rewards = defaultdict(list)
self.users_consumed_items = defaultdict(list)
self.users_consumed_items_rewards = defaultdict(list)
for i in tqdm(range(len(self.train_dataset.data))):
uid = int(self.train_dataset.data[i, 0])
item = int(self.train_dataset.data[i, 1])
reward = self.train_dataset.data[i, 2]
self.users_consumed_items[uid].append(item)
self.users_consumed_items_rewards[uid].append(reward)
self.items_consumed_users[item].append(uid)
self.items_consumed_users_rewards[item].append(reward)
mf_model = mf.PMF(
num_lat=self.num_lat, var=self.var, user_var=self.var_u, item_var=self.var_v
)
mf_model.fit(self.train_consumption_matrix)
for i in range(self.num_particles):
self.particles_us[i] = mf_model.users_weights
self.particles_vs[i] = mf_model.items_weights
def actions_estimate(self, candidate_actions):
"""actions_estimate.
Args:
candidate_actions: (user id, candidate_items)
Returns:
numpy.ndarray:
"""
uid = candidate_actions[0]
candidate_items = candidate_actions[1]
particle_idx = np.random.choice(self.particles_ids)
items_score = (
self.particles_us[particle_idx][uid]
@ self.particles_vs[particle_idx][candidate_items].T
)
return items_score, None
def update(self, observation, action, reward, info):
"""update.
Args:
observation:
action: (user id, item)
reward (float): reward
info:
"""
uid = action[0]
item = action[1]
additional_data = info
with threadpool_limits(limits=1, user_api="blas"):
updated_history = False
lambdas_u_i = np.empty(
shape=(self.num_particles, self.num_lat, self.num_lat)
)
zetas_u_i = np.empty(shape=(self.num_particles, self.num_lat))
mus_u_i = np.empty(shape=(self.num_particles, self.num_lat))
for i in range(self.num_particles):
v_j = self.particles_vs[i][self.users_consumed_items[uid]]
lambda_u_i = 1 / self.var * (v_j.T @ v_j) + 1 / self.particles_var_us[
i
] * np.eye(self.num_lat)
zeta_u_i = np.sum(
np.multiply(
v_j,
np.array(self.users_consumed_items_rewards[uid]).reshape(-1, 1),
),
axis=0,
)
lambdas_u_i[i] = lambda_u_i
zetas_u_i[i] = zeta_u_i
mus_u_i[i] = 1 / self.var * (np.linalg.inv(lambda_u_i) @ zeta_u_i)
weights = np.empty(self.num_particles)
for i in range(self.num_particles):
lambda_u_i, mu_u_i = lambdas_u_i[i], mus_u_i[i]
v_j = self.particles_vs[i][item, :]
cov = 1 / self.var + np.dot(np.dot(v_j.T, lambda_u_i), v_j)
w = scipy.stats.norm(np.dot(v_j.T, mu_u_i), cov).pdf(reward)
weights[i] = w
normalized_weights = _softmax(weights)
ds = np.random.choice(
range(self.num_particles), p=normalized_weights, size=self.num_particles
)
new_particles_us = np.empty(
shape=(self.num_particles, self.num_total_users, self.num_lat)
)
new_particles_vs = np.empty(
shape=(self.num_particles, self.num_total_items, self.num_lat)
)
new_particles_var_us = np.empty(shape=(self.num_particles))
new_particles_var_vs = np.empty(shape=(self.num_particles))
for i in range(self.num_particles):
d = ds[i]
new_particles_us[i] = self.particles_us[d]
new_particles_vs[i] = self.particles_vs[d]
new_particles_var_us[i] = self.particles_var_us[d]
new_particles_var_vs[i] = self.particles_var_vs[d]
if not updated_history:
self.users_consumed_items[uid].append(item)
self.users_consumed_items_rewards[uid].append(reward)
self.items_consumed_users[item].append(uid)
self.items_consumed_users_rewards[item].append(reward)
updated_history = True
for i in range(self.num_particles):
lambda_u_i, zeta_u_i = lambdas_u_i[i], zetas_u_i[i]
v_j = new_particles_vs[i][item, :]
lambda_u_i += 1 / self.var * (v_j @ v_j.T)
zeta_u_i += reward * v_j
inv_lambda_u_i = np.linalg.inv(lambda_u_i)
sampled_user_vector = np.random.multivariate_normal(
1 / self.var * (inv_lambda_u_i @ zeta_u_i), inv_lambda_u_i
)
new_particles_us[i][uid] = sampled_user_vector
u_i = new_particles_us[i][self.items_consumed_users[item], :]
lambda_v_i = 1 / self.var * (u_i.T @ u_i) + 1 / new_particles_var_vs[
i
] * np.eye(self.num_lat)
zeta = np.sum(
np.multiply(
u_i,
np.array(self.items_consumed_users_rewards[item]).reshape(
-1, 1
),
),
axis=0,
)
inv_lambda_v_i = np.linalg.inv(lambda_v_i)
item_sample_vector = np.random.multivariate_normal(
1 / self.var * (inv_lambda_v_i @ zeta), inv_lambda_v_i
)
new_particles_vs[i][item] = item_sample_vector
if not updated_history:
self.users_consumed_items[uid].append(item)
self.users_consumed_items_rewards[uid].append(reward)
self.items_consumed_users[item].append(uid)
self.items_consumed_users_rewards[item].append(reward)
updated_history = True
self.particles_us = new_particles_us
self.particles_vs = new_particles_vs
self.particles_var_us = new_particles_var_us
self.particles_var_vs = new_particles_var_vs