forked from Guangyi-Zhang/diverse-rule-set
-
Notifications
You must be signed in to change notification settings - Fork 0
/
decisionset.py
171 lines (146 loc) · 5.87 KB
/
decisionset.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
from itemset import Itemset, Rule, Transaction, prep_db
from div import GreedyDiv
from sampling import sample, sample_rn
from util import recall, precision, dispersion, printRules, obj
from sklearn import tree
from sklearn.metrics import accuracy_score, balanced_accuracy_score
from sklearn.model_selection import train_test_split
import numpy as np
from typing import Tuple
import operator
class DecisionSet(object):
def __init__(self, eps=0.01):
self.rules = []
self.labels = set()
self.recall_eps = eps
self.seq = True # True if use sol from 1st, or o.w. 2nd round
self.default = None
def set_default(self, label=None):
'''The most under-represented class'''
if label is not None:
self.default = label
return label
rc = recall(self.rules)
idx = np.argmin([rc[label] for label in Itemset.labels])
deft = Itemset.labels[idx]
self.default = deft
return self.default
def train(self, X, Y, max_k=100, nsamp=100, lamb=None, q='kl', mode=3, rerun=True,
min_recall_per_class=0.5):
print('##### START #####')
Itemset.clear_db()
prep_db(X, Y)
# Allow specify lamb to a certain number by users
if type(lamb) == str or lamb is None:
samp = self.sample_from_each_label(set(Itemset.labels), 100, set(), mode)
if lamb == 'max':
lamb = np.max([Rule.quality([r], metric=q) for r in samp])
elif lamb == 'mean':
lamb = np.mean([Rule.quality([r], metric=q) for r in samp])
else:
lamb = 0
print('lamb:', lamb)
greed = GreedyDiv([], lamb)
U_all = []
labels_samp = set(Itemset.labels)
while len(self) < max_k and len(labels_samp) > 0:
if mode == 0:
samps = []
for label in labels_samp:
_, samp = sample_rn(nsamp, label)
samp = [Rule(s, label) for s in list(samp)] # Very time-consuming
samps.extend(samp)
U = set(samps)
else:
covered = set([t for r in self.rules for t in r.trans()])
U = self.sample_from_each_label(labels_samp, nsamp, covered, mode)
print('nsamp (after):', len(U))
if len(U) == 0:
break
U_all.extend(U)
# Greedy
greed.update_univ(U)
r = greed.greedy_once()
# Termination criteria. Also check zero sampling above.
if self.enough(r):
# Include at least one rule per class, except default class.
labels_samp.remove(r.label)
print('remove label:', r.label)
else:
# Print quality vs. dispersion
q, d = obj(self.rules, lamb, sep=True)
qr, dr = obj(self.rules + [r], lamb, sep=True)
print('inc q vs. d: {}, {}'.format(qr-q, dr-d))
self.add(r)
if np.abs(recall(self.rules)[r.label] - 1.0) < 1e-8:
labels_samp.remove(r.label)
print('#{} '.format(len(self.rules)), end='')
printRules([r])
# Consecutive greedy over all sampels
if rerun:
greed.clear()
greed.update_univ(list(set(U_all)))
rules = greed.greedy(len(self.rules))
if obj(rules, lamb) > obj(self.rules, lamb):
print('Full greedy wins: {} > {}'.format(obj(rules, lamb), obj(self.rules, lamb)))
self.reset(rules)
default = self.set_default()
print('default:', default)
self.build()
print('precision: ', precision(self).items())
print('recall (coverage): ', recall(self.rules).items())
print('ave disp: ', dispersion(self.rules, average=True))
print('##### END #####')
def sample_from_each_label(self, labels_samp, nsamp, covered, mode):
# Sample rules from each label
samps = []
for label in labels_samp:
db0 = [db for l, db in Itemset.dbs.items() if l != label]
nsamp_, samp = sample(nsamp, db0, Itemset.dbs[label], covered, mode=mode)
if nsamp_ == 0:
continue
samp = [Rule(s, label) for s in list(samp)] # Very time-consuming
samps = samps + samp
return set(samps)
def add(self, rule: Rule):
self.rules.append(rule)
self.labels.add(rule.label)
def reset(self, rules: list):
self.seq = False
self.rules = rules
self.labels = {r.label for r in rules}
def enough(self, r: Rule) -> bool:
rc_cur = recall(self.rules)
rc_aft = recall(self.rules + [r])
if rc_aft[r.label] - rc_cur[r.label] <= self.recall_eps:
return True
return False
##################
# For prediction
##################
def build(self):
'''Run after adding all rules'''
self.rules = self._sort_rules(self.rules)
def _sort_rules(self, rules):
#vals = np.array([Rule.quality([r]) for r in rules])
vals = np.array([Rule.quality([r], metric='acc') for r in rules])
idx = np.argsort(-vals)
return [rules[i] for i in idx]
def predict_all(self, X: list) -> list:
return np.array([self.predict(x) for x in X])
def predict(self, x: Transaction) -> bool:
for r in self.rules:
if r.cover(x):
return r.label
return self.default
def predict_and_rule(self, x: Transaction) -> Tuple[bool, Rule]:
'''Return the rule with the highest discrim()'''
for r in self.rules:
if r.cover(x):
return r.label, r
return self.default, None
##################
# Util
##################
def __len__(self):
return len(self.rules)