-
Notifications
You must be signed in to change notification settings - Fork 0
/
auction.py
291 lines (239 loc) · 10.3 KB
/
auction.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
285
286
287
288
289
290
291
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from information import mi2eff
class base_auction(nn.Module):
def __init__(self, agents, c, r, datastream):
super(base_auction, self).__init__()
self.agents = agents
self.datastream = datastream
self.c = c
self.r = r
def run(self, data, logger):
pass
class recsys_auction(base_auction):
def __init__(self, agents, datastream, activation='sigmoid'):
super().__init__(agents, 0, 0, datastream)
self.users = datastream.users
self.uEmb = datastream.dataset.uEmb
self.vEmb = datastream.dataset.vEmb
self.sharpness = datastream.dataset.sharpness
if activation == 'sigmoid':
self.activation = self.sigmoid_pCTR
elif activation == 'id':
self.activation = self._pCTR
elif activation == 'binary':
self.activation = lambda x,y : torch.sign(self._pCTR(x,y))
def interact(self, pCTR):
interaction = 1.0 if torch.rand(1) < pCTR else 0.0
self.logger['pCTR'].append(float(pCTR))
self.logger['interaction'].append(int(interaction))
return interaction, pCTR
def _pCTR(self, rec, u):
return torch.sum(self.vEmb[rec] * self.uEmb[u])
def sigmoid_pCTR(self, rec, u):
return torch.sigmoid(self.sharpness*self._pCTR(rec,u))
def user_decision(self, data):
u,_ = data
u = u.squeeze()
self.logger['x'].append(int(u))
choice = self.users[u].choose()
self.logger['choices'].append(int(choice))
return u, choice
def query_choice_agent(self, u, choice):
rec = self.agents[choice].rec(u)
self.logger['recs'].append(int(rec))
self.logger['agents'][choice]['y_hat'].append(int(rec))
return rec
def query_all_agents(self, u):
all_recs = []
for a in self.agents:
rec = a.rec(u)
all_recs.append(rec)
self.logger['agents'][a.id]['y_hat'].append(int(rec))
return all_recs
def agent_logging(self):
for i,agent in enumerate(self.agents):
self.logger['agents'][agent.id]['reward'].append(int(agent.returns))
self.logger['agents'][agent.id]['wins'].append(int(agent.total_recs))
self.logger['agents'][agent.id]['uEmb'] = agent.U
self.logger['agents'][agent.id]['vEmb'] = agent.V
self.logger['agents'][agent.id]['records'] = agent.avg_r
self.logger['agents'][agent.id]['counts'] = agent.n
self.logger['agents'][agent.id]['random_tally'] = agent.random_tally
#log the ground truth here as well
if 'uEmb' not in self.logger.keys():
self.logger['uEmb'] = self.uEmb
self.logger['vEmb'] = self.vEmb
def update(self, interaction, pCTR, choice, u, rec):
self.agents[choice].update(u, rec, interaction)
self.users[u].update(choice, interaction)
self.agent_logging()
def run(self, data, logger):
u, choice = self.user_decision(data)
rec = self.query_choice_agent(u, choice)
self.update(*self.interact(self.activation(rec,u)), choice, u, rec)
class recsys_auction_with_baselines(recsys_auction):
def __init__(self, agents, datastream, activation='sigmoid'):
super().__init__(agents, datastream, activation=activation)
def run(self, data, logger):
u, choice = self.user_decision(data)
rec = self.query_all_agents(u)[choice]
self.update(*self.interact(self.activation(rec,u)), choice, u, rec)
class recsys_auction_debiased(recsys_auction_with_baselines):
def __init__(self, agents, datastream, activation='sigmoid'):
super().__init__(agents, datastream, activation=activation)
def update(self, interaction, pCTR, choice, u, rec):
rand_u = np.random.randint(len(self.users))
rand_rec = self.query_choice_agent(rand_u, choice)
interaction, _ = self.interact(self.activation(rand_rec,u))
self.agents[choice].update(rand_u, rand_rec, interaction)
self.users[u].update(choice, interaction)
self.agent_logging()
class classification_auction(base_auction):
def __init__(self, agents, datastream, c=1, r=2):
super().__init__(agents, c, r, datastream)
def _process_data(self, data):
x,y = data
#x = x.reshape(-1,1)
y = y.float()
return x,y
def score_models(self, data):
x,y = self._process_data(data)
self.logger['x'].append(x)
self.logger['y'].append(y)
y_hats = []
#compute predicted vals
scores = []
for i,a in enumerate(self.agents):
a.get_reward(-self.c)
y_hat = a.predict(x)
y_hats.append(y_hat)
scores.append(self.score(y,y_hat))
self.logger['scores'].append(scores)
return scores, y_hats, x, y
def system_correctness(self, scores):
correct_agents = []
for i,score in enumerate(scores):
if score == 1:
correct_agents.append(self.agents[i])
if correct_agents == []:
self.logger['agg-correct'].append(False)
correct_agents = self.agents
else:
self.logger['agg-correct'].append(True)
return correct_agents
def user_decision(self, scores):
correct_agents = self.system_correctness(scores)
wid = torch.randint(len(correct_agents), (1,))[0]
return correct_agents[wid]
def update_winner(self, winner, x, y):
self.logger['winner'].append(winner.id)
winner.get_reward(self.r)
winner.add_data(x,y)
winner._update(x,y)
winner.wins += 1
def update_agents(self, y_hats):
for i,agent in enumerate(self.agents):
self.logger['agents'][agent.id]['reward'].append(agent.reward)
self.logger['agents'][agent.id]['wins'].append(agent.wins)
self.logger['agents'][agent.id]['y_hat'].append(y_hats[i])
self.logger['agents'][agent.id][
'dataset_counts'] = agent.dataset_counts
def run(self, data, logger):
scores, y_hats, x, y = self.score_models(data)
winner = self.user_decision(scores)
self.update_winner(winner, x, y)
self.update_agents(y_hats)
def score(self, y, y_hat):
return 1 if y == y_hat else 0
class inefficient_classification_auction(classification_auction):
def __init__(self, agents, datastream, c=1, r=2, alpha=1, mi=None):
super().__init__(agents, datastream, c=c, r=r)
if mi == None:
self.alpha = alpha
else:
self.alpha = mi2eff(mi, len(agents), 1e-3)
def system_correctness(self, scores, wid):
corr_agent = super().system_correctness(scores)
self.logger['agg-correct'][-1] = self.logger['agg-correct'][-1] and \
wid in [a.id for a in corr_agent]
def user_decision(self, scores):
s_np = np.array(scores)
softmin = np.exp(self.alpha*s_np)/sum(np.exp(self.alpha*s_np))
wid = np.random.choice(range(len(s_np)), size=1, p=softmin)[0]
self.system_correctness(scores, wid)
return self.agents[wid]
class debiased_classification_auction(inefficient_classification_auction):
def __init__(self, agents, datastream, c=1, r=2, alpha=1, mi=None):
super().__init__(agents, datastream, c=c, r=r, alpha=alpha, mi=mi)
def update_winner(self, winner, x, y):
#need to debias the learning -- produce a random iid sample instead
idx = np.random.randint(len(self.datastream.dataset))
x,y = self.datastream.dataset[idx]
#x = torch.tensor(x)
x,y = self._process_data((
torch.tensor(x).unsqueeze(0),torch.tensor([y])))
super().update_winner(winner, x, y)
class regression_auction(inefficient_classification_auction):
def __init__(self, agents, datastream, c=1, r=2, alpha=1):
super().__init__(agents, datastream, c=c, r=r, alpha=alpha)
self.alpha = -alpha #alpha is negative for losses
def system_correctness(self, scores, wid):
self.logger['agg-correct'].append(scores[wid])
def score(self, y, y_hat): #for now just implement square error
return torch.sum((y-y_hat)**2)
class bidding_classification_auction(classification_auction):
def __init__(self, agents, c, r, datastream):
super().__init__(agents, c, r, datastream)
def run(self, data, logger):
#run one step of mechanism
x,y = data
x = x.reshape(-1,1)
y = y.float()
logger['x'].append(x)
logger['y'].append(y)
y_hats = []
bidders = set()
#compute predicted vals
scores = []
for i,a in enumerate(self.agents):
atarg = a.target(x)
logger['agents'][i]['bid'].append(atarg)
if atarg:
a.get_reward(-self.c)
bidders.add(a)
y_hat = a.predict(x)
y_hats.append(y_hat)
s = self.score(y,y_hat)
scores.append(self.score(y,y_hat))
logger['scores'].append(scores)
correct_agents = []
for i,score in enumerate(scores):
if score == 1 and self.agents[i] in bidders:
correct_agents.append(self.agents[i])
if correct_agents == []:
logger['agg-correct'].append(False)
correct_agents = list(bidders)
else:
logger['agg-correct'].append(True)
wid = -1
if len(correct_agents) > 0:
wid = torch.randint(len(correct_agents), (1,))[0]
winner = correct_agents[wid]
logger['winner'].append(winner.id)
winner.get_reward(self.r)
winner.add_data(x,y)
winner.wins += 1
else:
logger['winner'].append(-1)
logger['bidders'].append(bidders)
for i,agent in enumerate(self.agents):
net_reward = -self.c
net_reward += self.r if wid == agent.id else 0
agent.update(net_reward)
logger['agents'][agent.id]['reward'].append(agent.reward)
logger['agents'][agent.id]['wins'].append(agent.wins)
logger['agents'][agent.id]['y_hat'].append(y_hats[i])
logger['agents'][agent.id]['state'].append(agent.dist_bucket_est)