-
Notifications
You must be signed in to change notification settings - Fork 1
/
ruxg.py
581 lines (461 loc) · 22 KB
/
ruxg.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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
"""
@author: sibirbil
"""
import copy
import time
import warnings
import numpy as np
import cvxpy as cp
import gurobipy as gp
from gurobipy import GRB
from scipy.sparse import csr_matrix
from sklearn.base import ClassifierMixin, BaseEstimator
from sklearn.ensemble import GradientBoostingClassifier, GradientBoostingRegressor
from sklearn.utils.validation import check_is_fitted
from sklearn.exceptions import NotFittedError
from sklearn.tree import DecisionTreeClassifier
from auxClasses import SklearnEstimator, Coefficients, Clause, Rule
class RUXG(BaseEstimator, SklearnEstimator):
def __init__(self):
self.fitTime = 0
self.predictTime = 0
def _initialize(self):
self.fittedDTs = {}
self.initNumOfRules = 0
self.rules = {}
self.ruleInfo = {}
self.missedXvals = None
self.numOfMissed = None
self.rulesPerSample = None
self.ruleLengthPerSample = None
self.normConst = 1.0
# Coefficients (A, Abar & costs) are used as CSR sparse matrices
self._coeffs = Coefficients()
def _checkOptions(self):
if (type(self) == RUXClassifier):
if (self.trained_ensemble == None):
raise ValueError('One ensemble learning method (RF, ADA or GB) should be provided.')
try:
check_is_fitted(self.trained_ensemble)
except NotFittedError as e:
raise ValueError('Ensemble learning method should be fitted first to use in RUX.')
for treeno, fitTree in enumerate(self.trained_ensemble.estimators_):
if (type(self.trained_ensemble) == GradientBoostingClassifier or
type(self.trained_ensemble) == GradientBoostingRegressor):
self.initNumOfRules += fitTree[0].get_n_leaves()
self.fittedDTs[treeno] = fitTree[0]
else:
self.initNumOfRules += fitTree.get_n_leaves()
self.fittedDTs[treeno] = fitTree
def _cleanup(self):
self.labelToInteger = {}
self.integerToLabel= {}
self.fittedDTs = {}
self.rules = {}
self.ruleInfo = {}
self.missedXvals = None
self.numOfMissed = None
self.rulesPerSample = None
self.ruleLengthPerSample = None
self._coeffs._cleanup()
def _getRule(self, fitTree, nodeid):
if (fitTree.tree_.feature[0] == -2): # No rule case
return Rule()
left = fitTree.tree_.children_left
right = fitTree.tree_.children_right
threshold = fitTree.tree_.threshold
def recurse(left, right, child, returnRule=None):
if returnRule is None:
returnRule = Rule()
if child in left: # 'l'
parent = np.where(left == child)[0].item()
clause = Clause(feature=fitTree.tree_.feature[parent],
ub=threshold[parent])
else: # 'r'
parent = np.where(right == child)[0].item()
clause = Clause(feature=fitTree.tree_.feature[parent],
lb=threshold[parent])
returnRule.addClause(clause)
if parent == 0:
return returnRule
else:
return recurse(left, right, parent, returnRule)
retRule = recurse(left, right, nodeid)
return retRule
def _getMatrix(self, X, y, fitTree, treeno):
if (len(self._coeffs.cols) == 0):
col = 0
else:
col = max(self._coeffs.cols) + 1 # Next column
y_rules = fitTree.apply(X) # tells us which sample is in which leaf
for leafno in np.unique(y_rules):
covers = np.where(y_rules == leafno)[0]
leafYvals = y[covers] # y values of the samples in the leaf
uniqueLabels, counts = np.unique(leafYvals, return_counts=True)
label = uniqueLabels[np.argmax(counts)] # majority class in the leaf
labelVector = np.ones(self.K)*(-1/(self.K-1))
labelVector[self.labelToInteger[label]] = 1
fillAhat = np.dot(self.vecY[:, covers].T, labelVector)
self._coeffs.rows = np.hstack((self._coeffs.rows, covers))
self._coeffs.cols = np.hstack((self._coeffs.cols, np.ones(len(covers), dtype=np.int32)*col))
self._coeffs.yvals = np.hstack((self._coeffs.yvals, np.ones(len(covers), dtype=np.float64) * fillAhat))
if (self.rule_length_cost):
tempRule = self._getRule(fitTree, leafno)
cost = tempRule.length()
else:
cost = 1.0
self._coeffs.costs = np.append(self._coeffs.costs, cost)
self.ruleInfo[col] = (treeno, leafno, label)
col += 1
self.normConst = 1.0/np.max(self._coeffs.costs)
def _getMatrices(self, X, y):
if (type(self) == RUXClassifier):
for treeno, fitTree in enumerate(self.fittedDTs.values()):
self._getMatrix(X, y, fitTree, treeno)
def _preprocess(self, X, y):
classes, classCounts = np.unique(y, return_counts=True)
self.majorityClass = classes[np.argmax(classCounts)]
for i, c in enumerate(classes):
self.labelToInteger[c] = i
self.integerToLabel[i] = c
self.K = len(classes)
n = len(y)
self.vecY = np.ones((self.K, n))*(-1/(self.K-1))
for i, c in enumerate(y):
self.vecY[self.labelToInteger[c], i] = 1
def _fillRules(self, weights):
# The weights are scaled only for classification
if (np.max(weights > 1.0e-6)):
weights = weights/np.max(weights) # Scaled weights
selectedColumns = np.where(weights > self.threshold)[0] # selected columns
weightOrder = np.argsort(-weights[selectedColumns]) # ordered weights
orderedColumns = selectedColumns[weightOrder] # ordered indices
for i, col in enumerate(orderedColumns):
treeno, leafno, label = self.ruleInfo[col]
fitTree = self.fittedDTs[treeno]
if (fitTree.get_n_leaves()==1):
self.rules[i] = Rule(label=self.majorityClass,
clauses=[], weight=weights[col]) # no rule
else:
self.rules[i] = self._getRule(fitTree, leafno)
self.rules[i].label = label
self.rules[i].weight = weights[col]
self.rules[i]._cleanRule()
def _solvePrimal(self, y=None, ws0=[], vs0=[], groups=None):
# TODO: Pyomo
if(self.solver == 'glpk'):
return self._solvePrimalGLPK(y=y, ws0=ws0, vs0=vs0, groups=groups)
elif (self.solver == 'gurobi'):
return self._solvePrimalGurobi(y=y, ws0=ws0, vs0=vs0, groups=groups)
else:
raise ValueError('Solver {0} does not exist'.format(self.solver))
def _solvePrimalGLPK(self, y=None, ws0=[], vs0=[], groups=None):
Ahat = csr_matrix((self._coeffs.yvals, (self._coeffs.rows, self._coeffs.cols)), dtype=np.float64)
n, m = max(self._coeffs.rows)+1, max(self._coeffs.cols)+1
# Variables
vs = cp.Variable(n, nonneg=True)
ws = cp.Variable(m, nonneg=True)
if (len(vs0) > 0):
vs.value = vs0
if (len(ws0) > 0):
ws.value = np.zeros(m)
ws.value[:len(ws0)] = ws0
# Primal Model
primal = cp.Problem(cp.Minimize(np.ones(n) @ vs +
(self._coeffs.costs * self.pen_par * self.normConst) @ ws),
[(((self.K - 1.0)/self.K)*Ahat) @ ws + vs >= 1.0])
# Fairness constraints
if self.fair_metric==None:
primal.solve(solver=cp.GLPK, glpk={'msg_lev': 'GLP_MSG_OFF'})
if self.fair_metric=='dmc' or self.fair_metric=='odm':
for pair in groups:
setgroup1 = pair[0]
setgroup2 = pair[1]
amount1 = sum(setgroup1)
amount2 = sum(setgroup2)
if amount1==0 or amount2==0:
continue
else:
fairness_constraints = [(((self.K - 1.0)/self.K)*Ahat) @ ws + vs >= 1.0]
fairness_constraints.append((((1.0/amount1)*setgroup1) - ((1/amount2)*setgroup2)) @ vs <= self.fair_eps)
fairness_constraints.append((((1.0/amount2)*setgroup2) - ((1/amount1)*setgroup1)) @ vs <= self.fair_eps)
primal = cp.Problem(cp.Minimize(np.ones(n) @ vs +
(self._coeffs.costs * self.pen_par * self.normConst) @ ws),fairness_constraints)
primal.solve(solver=cp.GLPK, glpk={'msg_lev': 'GLP_MSG_OFF'})
betas = primal.constraints[0].dual_value
return ws.value, vs.value, betas
def _solvePrimalGurobi(self, y=None, ws0=[], vs0=[], groups=None):
Ahat = csr_matrix((self._coeffs.yvals, (self._coeffs.rows, self._coeffs.cols)), dtype=np.float64)
n, m = max(self._coeffs.rows)+1, max(self._coeffs.cols)+1
# Primal Model
modprimal = gp.Model('RUXG Primal')
modprimal.setParam('OutputFlag', False)
# Variables
vs = modprimal.addMVar(shape=int(n), name='vs')
ws = modprimal.addMVar(shape=int(m), name='ws')
if (len(vs0) > 0):
vs.setAttr('Start', vs0)
modprimal.update()
if (len(ws0) > 0):
tempws = np.zeros(m)
tempws[:len(ws0)] = ws0
ws.setAttr('Start', tempws)
modprimal.update()
# Objective
modprimal.setObjective(np.ones(n) @ vs +
(self._coeffs.costs * self.pen_par * self.normConst) @ ws, GRB.MINIMIZE)
# Constraints
modprimal.addConstr((((self.K - 1.0)/self.K)*Ahat) @ ws + vs >= 1.0, name='Ahat Constraints')
# Fairness constraints
if self.fair_metric==None:
modprimal.optimize()
elif self.fair_metric == 'dmc' or self.fair_metric == 'odm' or self.fair_metric=='EqOpp':
for pair in groups: # groups is a pair of groups
setgroup1 = pair[0]
setgroup2 = pair[1]
amount1 = sum(setgroup1)
amount2 = sum(setgroup2)
if amount1==0 or amount2==0:
continue
else:
modprimal.addConstr((((1.0/amount1)*setgroup1) - ((1/amount2)*setgroup2)) @ vs <= self.fair_eps, name='Fairness constraints 1')
modprimal.addConstr((((1.0/amount2)*setgroup2) - ((1/amount1)*setgroup1)) @ vs <= self.fair_eps, name='Fairness constraints 2')
modprimal.optimize()
betas = np.array(modprimal.getAttr(GRB.Attr.Pi)[:n])
return ws.X, vs.X, betas
def print_rules(self, feature_names=None, indices=[]):
if (len(indices) == 0):
indices = self.rules.keys()
elif (np.max(indices) > len(self.rules)):
warnings.warn('\n There are only {0} rules'.format(len(self.rules)))
return
for indx in indices:
rule = self.rules[indx]
print('RULE %d:' % (indx))
if (rule.length() == 0):
print('==> No Rule: Set Majority Class')
else:
rule.printRule(feature_names)
print('Class: %.0f' % rule.label)
print('Scaled rule weight: %.4f\n' % rule.weight)
# Gets the rules used for prediction for each instance
def get_instance_to_rule_dicts(self, IDs, X_test, rule_w_thresh=0):
x_id_to_rule_ids_dict = {}
x_id_to_rule_num_dict = {}
rule_ids = self.rules.keys()
for x0_id in IDs:
rule_num = 0
x0 = X_test.loc[x0_id].to_numpy()
sumClassWeights = np.zeros(self.K)
x_id_to_rule_ids_dict[x0_id] = []
for rule_id in rule_ids:
rule = self.rules[rule_id]
if rule.checkRule(x0) and rule.weight > rule_w_thresh:
rule_num += 1
x_id_to_rule_ids_dict[x0_id].append(rule_id)
x_id_to_rule_num_dict[x0_id] = rule_num
return x_id_to_rule_ids_dict, x_id_to_rule_num_dict
def print_rules_for_instances(self, IDs, x_id_to_rule_ids_dict, feature_names=None):
for x0_id in IDs:
print('Rules for the instance:\n')
rule_ids = x_id_to_rule_ids_dict[x0_id]
self.print_rules(indices=rule_ids, feature_names=feature_names)
print('\n \n')
def print_weights(self, indices=[]):
if (len(indices) == 0):
indices = self.rules.keys()
elif (np.max(indices) > len(self.rules)):
warnings.warn('\n There are only {0} rules'.format(len(self.rules)))
return
for indx in indices:
rule = self.rules[indx]
print('RULE %d:' % (indx))
print('Value: %.0f' % rule.label)
print('Scaled rule weight: %.4f\n' % rule.weight)
def get_weights(self, indices=[]):
if (len(indices) == 0):
indices = self.rules.keys()
elif (np.max(indices) > len(self.rules)):
warnings.warn('\n There are only {0} rules'.format(len(self.rules)))
return None
return [self.rules[indx].weight for indx in indices]
def get_avg_rule_length(self):
return np.mean([rule.length() for rule in self.rules.values()])
def get_num_of_rules(self):
return len(self.rules)
def get_init_num_of_rules(self):
return self.initNumOfRules
def get_num_of_missed(self):
return self.numOfMissed
def get_avg_num_rules_per_sample(self):
return np.mean(self.rulesPerSample)
def get_avg_rule_length_per_sample(self):
return np.mean(self.ruleLengthPerSample)
def get_fit_time(self):
return self.fitTime
def get_predict_time(self):
return self.predictTime
def predict(self, X, indices=[]):
if (self.fittedDTs == {}):
raise ValueError('You need to fit the RUG model first')
if (len(indices) == 0):
indices = self.rules.keys()
elif (np.max(indices) > len(self.rules)):
warnings.warn('\n There are only {0} rules'.format(len(self.rules)))
return None
self.missedXvals = []
self.numOfMissed = 0
self.rulesPerSample = np.zeros(len(X))
self.ruleLengthPerSample = np.zeros(len(X))
startTime = time.time()
returnPrediction = []
for sindx, x0 in enumerate(X):
sumClassWeights = np.zeros(self.K)
rule_lengths = []
for indx in indices:
rule = self.rules[indx]
if(rule.checkRule(x0)):
lab2int = self.labelToInteger[rule.label]
sumClassWeights[lab2int] += rule.weight
self.rulesPerSample[sindx] += 1.0
rule_lengths.append(rule.length())
if len(rule_lengths) > 0:
self.ruleLengthPerSample[sindx] = np.mean(rule_lengths)
if (np.sum(sumClassWeights) == 0):
# Unclassified test sample
self.numOfMissed += 1
self.missedXvals.append(x0)
# Assigned to a class with the initial DT
getClass = self.fittedDTs[0].predict(x0.reshape(1, -1))[0]
returnPrediction.append(getClass)
else:
sel_label_indx = np.argmax(sumClassWeights)
int2lab = self.integerToLabel[sel_label_indx]
returnPrediction.append(int2lab)
endTime = time.time()
self.predictTime = endTime - startTime
return returnPrediction
class RUXClassifier(RUXG, ClassifierMixin):
def __init__(self, trained_ensemble=None, pen_par=1.0,
threshold=1.0e-6, rule_length_cost=False,
solver='gurobi', fair_eps=1.0, fair_metric=None, random_state=None):
self._initialize()
self.trained_ensemble = trained_ensemble
self.pen_par = pen_par
self.threshold = threshold
self.solver = solver
self.fair_eps = fair_eps
self.fair_metric = fair_metric
self.random_state = random_state
self.rng_ = np.random.default_rng(self.random_state)
self.K = None # number of classes
self.labelToInteger = {} # mapping classes to integers
self.integerToLabel= {} # mapping integers to classes
self.vecY = None
self.majorityClass = None
self.rule_length_cost = rule_length_cost
# Classifier type
self._checkOptions()
def fit(self, X, y, groups=None):
if (len(self._coeffs.cols) != 0):
self._cleanup()
if (self.fair_metric == None):
groups = None
elif(groups == None):
warnings.warn('Groups should be provided. Setting fairness metric to None.')
self.fair_metric = None
startTime = time.time()
self._preprocess(X, y)
self._getMatrices(X, y)
ws = self._solvePrimal(groups=groups)[0]
self._fillRules(ws)
endTime = time.time()
self.fitTime = endTime - startTime
return self
class RUGClassifier(RUXG, ClassifierMixin):
'''
Parameters
----------
pen_par:
'''
def __init__(self, pen_par=1.0, threshold=1.0e-6,
max_depth=None, max_RMP_calls=30, rule_length_cost=False,
solver='gurobi', fair_eps=1.0, fair_metric=None, random_state=None):
self._initialize()
self.pen_par = pen_par
self.threshold = threshold
self.solver = solver
self.fair_eps = fair_eps
self.fair_metric = fair_metric
self.random_state = random_state
self.rng_ = np.random.default_rng(self.random_state)
self.K = None # number of classes
self.labelToInteger = {} # mapping classes to integers
self.integerToLabel= {} # mapping integers to classes
self.vecY = None
self.majorityClass = None
self.max_depth = max_depth
self.max_RMP_calls = max_RMP_calls
self.rule_length_cost = rule_length_cost
def _PSPDT(self, X, y, fitTree, treeno, betas):
n, col = len(X), max(self._coeffs.cols)+1
y_rules = fitTree.apply(X) # tells us which sample is in which leaf
noImprovement = True
for leafno in np.unique(y_rules):
covers = np.where(y_rules == leafno)[0]
# Prepare to check the reduced cost
aijhat = np.zeros(n)
leafYvals = y[covers] # y values of the samples in the leaf
uniqueLabels, counts = np.unique(leafYvals, return_counts=True)
label = uniqueLabels[np.argmax(counts)] # majority class in the leaf
labelVector = np.ones(self.K)*(-1.0/(self.K-1))
labelVector[self.labelToInteger[label]] = 1
fillAhat = np.dot(self.vecY[:, covers].T, labelVector)
aijhat[covers] = fillAhat
if (self.rule_length_cost):
tempRule = self._getRule(fitTree, leafno)
cost = tempRule.length()
else:
cost = 1.0
red_cost = np.dot((((self.K-1.0)/self.K)*aijhat), betas) - (cost * self.pen_par * self.normConst)
if (red_cost > 0): # only columns with proper reduced costs are added
self._coeffs.rows = np.hstack((self._coeffs.rows, covers))
self._coeffs.cols = np.hstack((self._coeffs.cols, np.ones(len(covers), dtype=np.int32)*col))
self._coeffs.yvals = np.hstack((self._coeffs.yvals, np.ones(len(covers), dtype=np.float64) * fillAhat))
self._coeffs.costs = np.append(self._coeffs.costs, cost)
self.ruleInfo[col] = (treeno, leafno, label)
col += 1
noImprovement = False
return noImprovement
def fit(self, X, y, groups=None):
if (len(self._coeffs.cols) != 0):
self._cleanup()
if (self.fair_metric == None):
groups = None
elif(groups == None):
warnings.warn('Groups should be provided. Setting fairness metric to None.')
self.fair_metric = None
startTime = time.time()
treeno = 0
DT = DecisionTreeClassifier(max_depth=self.max_depth,
random_state=self.rng_.integers(np.iinfo(np.int16).max))
fitTree = DT.fit(X, y)
self.fittedDTs[treeno] = copy.deepcopy(fitTree)
self._preprocess(X, y)
self._getMatrix(X, y, fitTree, treeno)
ws, vs, betas = self._solvePrimal(groups=groups)
# Column generation
for _ in range(self.max_RMP_calls):
treeno += 1
DT = DecisionTreeClassifier(max_depth=self.max_depth,
random_state=self.rng_.integers(np.iinfo(np.int16).max))
fitTree = DT.fit(X, y, sample_weight=betas) # use duals as weights
self.fittedDTs[treeno] = copy.deepcopy(fitTree)
noImprovement = self._PSPDT(X, y, fitTree, treeno, betas)
if (noImprovement):
break
ws, vs, betas = self._solvePrimal(ws0=ws, vs0=vs, groups=groups)
self._fillRules(ws)
endTime = time.time()
self.fitTime = endTime - startTime
return self