-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathws_ranking.py
238 lines (189 loc) · 8.06 KB
/
ws_ranking.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
from ws_lib import *
from mallows import *
import numpy as np
import logging
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger(__name__)
class WeakSupRanking:
def __init__(self, r_utils):
self.thetas = None
self.r_utils = r_utils
def train(self, conf, L, numLFs=None):
"""
Parameters
----------
conf: configuration dictionary, keys: 'train_method', 'inference_rule'
L: labels, dim (n, m) where n is the number of examples and m is the number of LFs (numLFs)
numLFs: the number of labeling functions
Returns
-------
"""
n = len(L) # the number of examples
if numLFs is None:
m = len(L[0])
else:
m = numLFs
d = len(L[0][0])
if conf['train_method'] == 'triplet' or conf['train_method'] == 'triplet_opt':
# TODO: handle the case when num LF is not multiple of 3
expected_dists = np.zeros(m)
i = 0
order, D = self.order_LFs_on_correlation(L,m)
#order = np.arange(m)
#D = None
# solve 3 variable systems iteratively
while i < m-2:
l1, l2, l3 = order[i], order[i+1], order[i+2]
ac3 = self.solve_triplet(L, l1, l2, l3, D)
expected_dists[l1] = ac3[0]
expected_dists[l2] = ac3[1]
expected_dists[l3] = ac3[2]
i += 3
if conf['train_method'] == 'triplet':
self.thetas = 1.0 / np.array(expected_dists)
elif conf['train_method'] == 'triplet_opt':
mlw = Mallows(self.r_utils, 1.0)
logger.info("expected_dists {}".format(expected_dists))
self.thetas = np.array([mlw.estimate_theta(d, expected_dists[i])
for i in range(len(expected_dists))])
self.thetas = self.thetas.clip(1e-1, 100)
elif (conf['train_method'] == 'median_triplet') or (conf['train_method'] == 'median_triplet_opt'):
expected_dists_dict = {}
for lf_idx in range(m):
expected_dists_dict[lf_idx] = []
order, D = self.order_LFs_on_correlation(L,m)
for i in range(m):
l1 = order[i]
for j in range(i+1, m):
l2 = order[j]
for k in range(j+1, m):
l3 = order[k]
ac3 = self.solve_triplet(L, l1, l2, l3, D)
expected_dists_dict[l1].append(ac3[0])
expected_dists_dict[l2].append(ac3[1])
expected_dists_dict[l3].append(ac3[2])
expected_dists = np.zeros(m)
for lf_idx in range(m):
expected_dists[lf_idx] = np.median(expected_dists_dict[lf_idx])
if conf['train_method'] == 'median_triplet':
self.thetas = 1.0 / np.array(expected_dists)
elif conf['train_method'] == 'median_triplet_opt':
mlw = Mallows(self.r_utils, 1.0)
logger.info("expected_dists {}".format(expected_dists))
self.thetas = np.array([mlw.estimate_theta(d, expected_dists[i])
for i in range(len(expected_dists))])
self.thetas = self.thetas.clip(1e-1, 100)
def mean_distance(self, L, l1, l2, normalize=False):
"""
Calculate mean distance of two distance, kendall tau distance is used as distance
Parameters
----------
L: labels, dim (n, m) where n is the number of examples and m is the number of LFs (numLFs)
l1: label function 1
l2: label function 2
normalize
Returns
-------
"""
n = len(L)
mu = np.mean([self.r_utils.kendall_tau_distance(L[i][l1], L[i][l2], normalize=normalize)
for i in range(n)],axis=0)
logger.info("mu {}".format(mu))
return mu
def order_LFs_on_correlation(self,L,numLFs=None):
"""
Parameters
----------
L: labels, dim (n, m) where n is the number of examples and m is the number of LFs (numLFs)
Returns
-------
order: order of label functions based on their max distance
D: the distance matrix of label functions
"""
if(numLFs):
m = numLFs
else:
m = len(L[0]) # the number of label functions
dists = [] # the list of triplets (i, j, distance(i,j))
D = np.zeros((m, m)) # the mean distance between label functions
# fill out the distance matrix
for i in range(m):
for j in range(i):
d_ij = self.mean_distance(L, i, j)
dists.append((i, j, d_ij))
D[i][j] = d_ij
D[j][i] = d_ij
dists = sorted(dists, key=lambda x: x[2], reverse=True)
# order of label functions based on their max distance
order = []
for i, j, d in dists:
if not i in order:
order.append(i)
if not j in order:
order.append(j)
return order, D
def solve_triplet(self, L, l1, l2, l3, D=None):
"""
Parameters
----------
L: labels, dim (n, m) where n is the number of examples and m is the number of LFs (numLFs)
l1: label function 1
l2: label function 2
l3: label function 3
D: distance matrix of label functions
Returns
-------
"""
if D is None:
mu_12 = self.mean_distance(L, l1, l2)
mu_23 = self.mean_distance(L, l2, l3)
mu_31 = self.mean_distance(L, l3, l1)
else:
mu_12 = D[l1][l2]
mu_23 = D[l2][l3]
mu_31 = D[l3][l1]
ac3 = solve_3_var_system_sum(mu_12, mu_23, mu_31)
return ac3
def infer_ranking(self,conf,L,numLFs=None,lst_D=None):
"""
Parameters
----------
conf: configuration dictionary, keys: 'train_method', 'inference_rule'
L: labels, dim (n, m) where n is the number of examples and m is the number of LFs (numLFs)
numLFs
lst_D
Returns
-------
"""
if numLFs is None:
numLFs = len(L[0])
k = numLFs
Y_tilde = None
n = len(L)
# label inference based on kemeny rule
if conf['inference_rule'] =='kemeny':
if lst_D is None:
Y_tilde = [self.r_utils.kemeny(L[i][:k]) for i in range(n)]
else:
Y_tilde = [self.r_utils.kemeny(L[i][:k],lst_D[i][:k, :k]) for i in range(n)]
# label inference based on weighted kemeny rule
elif conf['inference_rule'] == 'weighted_kemeny':
if lst_D is None:
Y_tilde = [self.r_utils.weighted_kemeny(L[i][:k], self.thetas[:k]) for i in range(n)]
else:
Y_tilde = [self.r_utils.weighted_kemeny(L[i][:k], self.thetas[:k], lst_D[i][:k,:k])
for i in range(n)]
# label inference based on pairwise majority
elif conf['inference_rule'] == 'pairwise_majority':
Y_tilde = [self.r_utils.pair_wise_majority(L[i][:k], weights=None) for i in range(n)]
# label inference based on weighted pairwise majority
elif conf['inference_rule'] == 'weighted_pairwise_majority':
Y_tilde = [self.r_utils.pair_wise_majority(L[i][:k], weights=self.thetas[:k]) for i in range(n)]
# label inference based on position estimation
elif conf['inference_rule'] == 'position_estimation':
Y_tilde = [self.r_utils.pos_est_majority(L[i][:k], weights=None) for i in range(n)]
# label inference based on weighted position estimation
elif conf['inference_rule']=='weighted_position_estimation':
Y_tilde = [self.r_utils.pos_est_majority(L[i][:k], weights=self.thetas[:k]) for i in range(n)]
h = [y.mask_items([]) for y in Y_tilde]
return Y_tilde