/
phcb.py
124 lines (107 loc) · 3.92 KB
/
phcb.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
import numpy as np
import math
from config import cfg
from collections import defaultdict
from naive_linucb import NaiveLinUCB
class ArmStruct():
def __init__(self,arm,depth):
self.arm = arm
self.gids = arm.gids
self.depth = depth # only when arm is a tree node
self.itemclick = defaultdict(bool)
self.feedback = defaultdict(float)
self.vv = defaultdict(int)
def expand(self):
if (sum(self.vv.values()))<cfg.activate_num*np.log(self.depth):
return False
if (sum(self.feedback.values())/sum(self.vv.values()))<cfg.activate_prob*np.log(self.depth):
return False
return True
class LinUCBUserStruct():
def __init__(self,uid,arms,dim,init,alpha):
self.uid = uid
self.arms = arms
self.dim = dim
self.A = np.identity(n=self.dim)
self.Ainv = np.linalg.inv(self.A)
self.b = np.zeros((self.dim))
if init!='zero':
self.theta = np.random.rand(self.dim)
else:
self.theta = np.zeros(self.dim)
self.alpha = alpha # store the new alpha calcuated in each iteratioin
if not cfg.random_choice:
self.linucb = NaiveLinUCB(cfg.linucb_para)
def getProb(self,fv):
if self.alpha==-1:
raise AssertionError
mean=np.dot(self.theta.T,fv)
var=np.sqrt(np.dot(np.dot(fv.T,self.Ainv),fv))
pta=mean+self.alpha*var
return pta
def getInv(self, old_Minv, nfv):
# new_M=old_M+nfv*nfv'
# try to get the inverse of new_M
tmp_a=np.dot(np.outer(np.dot(old_Minv,nfv),nfv),old_Minv)
tmp_b=1+np.dot(np.dot(nfv.T,old_Minv),nfv)
new_Minv=old_Minv-tmp_a/tmp_b
return new_Minv
def updateParameters(self, a_fv, reward):
self.A+=np.outer(a_fv, a_fv)
self.b+=a_fv*reward
self.Ainv=self.getInv(self.Ainv,a_fv)
self.theta=np.dot(self.Ainv, self.b)
class pHCB():
def __init__(self,para,init='zero',root=None):
self.init=init
self.users={}
try:
self.alpha=para['alpha']
except:
self.alpha=-1
self.root = root
def decide(self,uid,root=None):
try:
user=self.users[uid]
except:
arms = []
root = self.root
for node in root.children:
arms.append(ArmStruct(node,2))
dim = root.emb.shape[0]
self.users[uid]=LinUCBUserStruct(uid,arms,dim,"zero",self.alpha)
user=self.users[uid]
aid = None
max_r = float('-inf')
if len(user.arms)<=cfg.poolsize:
aids = list(range(len(user.arms)))
else:
aids = np.random.choice(len(user.arms),cfg.poolsize,replace=False)
for index in aids:
arm = user.arms[index]
depth = arm.depth
reward = user.getProb(arm.arm.emb)*depth
if reward>max_r:
aid = index
max_r = reward
arm_picker = user.arms[aid]
return arm_picker,aid
def updateParameters(self, picked_arm, reward, uid):
self.users[uid].updateParameters(picked_arm.arm.emb,reward)
def update(self,uid,aid,arm_picker,item,feedback):
gid = item.gid
arm_picker.feedback[gid] += feedback
arm_picker.vv[gid] += 1
arm_picker.itemclick[gid] = True
user = self.users[uid]
self.updateParameters(arm_picker,feedback,uid)
if arm_picker.expand() and arm_picker.arm.is_leaf==False:
depth = arm_picker.depth+1
user.arms.pop(aid)
for node in arm_picker.arm.children:
arm = ArmStruct(node,depth)
for gid in arm.gids:
arm.itemclick[gid]=arm_picker.itemclick[gid]
arm.feedback[gid]=arm_picker.feedback[gid]
arm.vv[gid]=arm_picker.vv[gid]
user.arms.append(arm)