-
Notifications
You must be signed in to change notification settings - Fork 1
/
synthetic_experiment.py
188 lines (159 loc) · 9.27 KB
/
synthetic_experiment.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
#!/usr/bin/env python
import numpy as np
import pandas as pd
from siscm import SISCM
from helper import *
class SyntheticExperiment:
def __init__(self, n_classes, n_features, n_groups, size_max, size_min, seed=44):
self.rng = np.random.default_rng(seed)
self.n_classes = n_classes
self.n_features = n_features
self.n_groups = n_groups
#create synthetic experts
self.group_size = self.rng.integers(size_min,size_max,size=self.n_groups)
self.n_experts = sum(self.group_size)
print("size of expert groups: ", self.group_size)
print("Total number of experts: ", self.n_experts)
self.create_data_model()
#retrieves categorical probabilities for individuals
def get_proba(weights, datapoint):
exp_wx = np.exp(np.dot(datapoint, weights))
return exp_wx/sum(exp_wx)
def create_data_model(self):
#sample discrete choice model weights
H_weights = self.rng.random( (self.n_experts, self.n_features, self.n_classes))
#array of functions that return group probabilities
H_marginal_proba_func = [(lambda d,w=w: SyntheticExperiment.get_proba(w, d)) for w in H_weights]
# list of sets with group members (index)
group_members = [ set(list(range(sum(self.group_size[:g]), sum(self.group_size[:g+1])))) for g in range(self.n_groups) ]
self.siscm_true = SISCM("Real", self.n_classes, H_marginal_proba_func, group_members)
def create_synthetic_data(self,N_training, N_test):
#create dummy data
data_train = self.rng.random((N_training, self.n_features))
data_test = self.rng.random((N_test, self.n_features))
#sample dummy labels
label_train = self.siscm_true.predict(data_train, 1)
label_test = self.siscm_true.predict(data_test, 1)
return (data_train, data_test, label_train, label_test)
#run synthetic experiment for values in sparsity list and #training points list
def run_experiment(self, T, N_test, sparsity_proba_list, N_train_list):
N_len = len(N_train_list)
s_len = len(sparsity_proba_list)
#empty matrices to store result
scores_real = np.full((T,s_len,N_len), np.nan)
scores_trained = np.full((T,s_len,N_len), np.nan)
scores_naive = np.full((T,s_len,N_len), np.nan)
scores_groups = np.full((T,s_len,N_len), np.nan)
ratio_inedges = np.full((T,s_len,N_len), np.nan)
# T experiment rounds
for t in range(T):
#create train and test data from true model M(Psi*)
full_data_train, data_test, full_labels_train, label_test = self.create_synthetic_data(max(N_train_list), N_test)
for s, sparsity_proba in enumerate(sparsity_proba_list):
max_N = max(N_train_list)
#sample experts' labels to drop according to sparsity probability
missing_inds = [self.rng.choice(a=range(self.n_experts), size = ( int(sparsity_proba*self.n_experts)), replace=False) for x in range(max_N)]
missing_inds = np.vstack(missing_inds)
#print(missing_inds.shape)
#leave at least two labels per datapoint
#sample 2 expert labels to keep
minimum_inds = [self.rng.choice(a=range(self.n_experts), size = (2), replace = False) for x in range(max_N)]
minimum_inds = np.vstack(minimum_inds)
#copy labels to new array
sparse_labels = np.copy(full_labels_train)
minimum_labels = sparse_labels[ np.arange(max_N)[:, np.newaxis], minimum_inds]
#remove sampled experts' labels, but keep the 2 sampled labels
sparse_labels[np.arange(max_N)[:,np.newaxis], missing_inds] = -999
sparse_labels[np.arange(max_N)[:, np.newaxis], minimum_inds] = minimum_labels
#sample expert's to predict during test
test_inds = self.rng.integers(self.n_experts, size = N_test)
#vary the number of training data
for n, N_train in enumerate(N_train_list):
print(t)
print(s, " ", sparsity_proba)
print(n, " ", N_train)
#remove excess datapoints
data_training = full_data_train[0:N_train]
label_training = sparse_labels[0:N_train]
#use true model M(Psi*) to find the ratio of edges inside the true groups
ratio_inedges[t,s,n] = self.siscm_true.analyze_PCS_graph( data_training, label_training)
if ratio_inedges[t,s,n]==1.0: continue
#create and train SI-SCM M(Psi)
siscm_psi = SISCM("SISCM_M(Psi)", self.n_classes, self.siscm_true.get_proba_function(),n_samples=500)
siscm_psi.fit( data_training, label_training)
#create SI-SCM M(H)
siscm_H = SISCM("SISCM_M(H)", self.n_classes, self.siscm_true.get_proba_function(), siscm_H= True)
#compare greedy algorithm groups to real groups
scores_groups[t,s,n] = compare_groups(siscm_psi, self.siscm_true)
#compare performace of the models for counterfactual predictions in the group
scores_real[t,s,n] = self.siscm_true.score_counterfactuals_rand(data_test, label_test, test_inds, label_test[range(N_test),test_inds])
scores_trained[t,s,n]= siscm_psi.score_counterfactuals_rand(data_test, label_test, test_inds, label_test[range(N_test),test_inds])
scores_naive[t,s,n] = siscm_H.score_counterfactuals_rand(data_test, label_test, test_inds, label_test[range(N_test),test_inds])
#print each rounds result
print("Log Likelihood for this round:")
print("Real: ", scores_real[t,s,n])
print("Trained: ", scores_trained[t,s,n])
print("Naive: ", scores_naive[t,s,n])
print("ARI: ", scores_groups[t,s,n])
print("Rate: ", ratio_inedges[t,s,n])
#compute mean and standard deviation
mean_score_real = np.mean(scores_real, axis=0)
std_score_real = np.std(scores_real, axis=0)
mean_score_trained = np.mean(scores_trained, axis=0)
std_score_trained = np.std(scores_trained, axis=0)
mean_score_naive = np.mean(scores_naive, axis=0)
std_score_naive = np.std(scores_naive, axis=0)
mean_score_groups = np.mean(scores_groups, axis=0)
std_score_groups = np.std(scores_groups, axis=0)
mean_rate_inedge = np.mean(ratio_inedges, axis=0)
std_rate_inedge = np.std(ratio_inedges, axis=0)
print()
print("Mean Score for Group of Observed Experts")
print("Score Naive CF: ", mean_score_naive, "+-", std_score_naive)
print("Score CF: ", mean_score_trained, "+-", std_score_trained)
print("Score Real CF: ", mean_score_real, "+-", std_score_real)
#save results to file
df_mean_real = pd.DataFrame(mean_score_real, columns = N_train_list, index= sparsity_proba_list)
df_mean_real.to_csv("results_synthetic/mean_real.csv")
df_std_real = pd.DataFrame(std_score_real, columns = N_train_list, index= sparsity_proba_list)
df_std_real.to_csv("results_synthetic/std_real.csv")
df_mean_trained = pd.DataFrame(mean_score_trained, columns = N_train_list, index= sparsity_proba_list)
df_mean_trained.to_csv("results_synthetic/mean_trained.csv")
df_std_trained = pd.DataFrame(std_score_trained, columns = N_train_list, index= sparsity_proba_list)
df_std_trained.to_csv("results_synthetic/std_trained.csv")
df_mean_naive = pd.DataFrame(mean_score_naive, columns = N_train_list, index= sparsity_proba_list)
df_mean_naive.to_csv("results_synthetic/mean_naive.csv")
df_std_naive = pd.DataFrame(std_score_naive, columns = N_train_list, index= sparsity_proba_list)
df_std_naive.to_csv("results_synthetic/std_naive.csv")
df_mean_groups = pd.DataFrame(mean_score_groups, columns = N_train_list, index= sparsity_proba_list)
df_mean_groups.to_csv("results_synthetic/mean_groups.csv")
df_std_groups = pd.DataFrame(std_score_groups, columns = N_train_list, index= sparsity_proba_list)
df_std_groups.to_csv("results_synthetic/std_groups.csv")
df_mean_inedge = pd.DataFrame(mean_rate_inedge, columns = N_train_list, index= sparsity_proba_list)
df_mean_inedge.to_csv("results_synthetic/mean_inedge.csv")
df_std_inedge = pd.DataFrame(std_rate_inedge, columns = N_train_list, index= sparsity_proba_list)
df_std_inedge.to_csv("results_synthetic/std_inedge.csv")
#scm_model.save()
def main():
seed = 44
#data parameters
n_classes = 5
n_features = 20
#true groups parameters
n_groups = 5
size_max = 15 #maximum number of experts per group
size_min = 5 #minimum number of experts per group
#create synthetic experiment with parameters
exp = SyntheticExperiment(n_classes, n_features, n_groups, size_max, size_min, seed)
#number of test datapoints
N_test=1000
#number of experiment rounds
T= 5
#list of different amounts of training data
N_train_list = [10, 20, 50, 75, 100, 150, 200, 300, 400]
#list of different sparsity percentage/ratio
sparsity_proba_list = [0.1, 0.3, 0.6, 0.8, 0.95]
#run experiment for set parameter lists
exp.run_experiment(T, N_test, sparsity_proba_list, N_train_list)
if __name__ == "__main__":
main()