def findClosestObservableSample(observable_samples, dataset_obj, factual_sample, norm_type):
closest_observable_sample = {'sample': {}, 'distance': np.infty, 'norm_type': None} # in case no observables are found.
for observable_sample_index, observable_sample in observable_samples.items():
# observable_sample['y'] = True
if observable_sample['y'] == factual_sample['y']: # make sure the cf flips the prediction
continue
# Only compare against those observable samples that DO NOT differ with the
# factual sample in non-actionable features!
violating_actionable_attributes = False
for attr_name_kurz in dataset_obj.getInputAttributeNames('kurz'):
attr_obj = dataset_obj.attributes_kurz[attr_name_kurz]
if attr_obj.actionability == 'none' and factual_sample[attr_name_kurz] != observable_sample[attr_name_kurz]:
violating_actionable_attributes = True
elif attr_obj.actionability == 'same-or-increase' and factual_sample[attr_name_kurz] > observable_sample[attr_name_kurz]:
violating_actionable_attributes = True
elif attr_obj.actionability == 'same-or-decrease' and factual_sample[attr_name_kurz] < observable_sample[attr_name_kurz]:
violating_actionable_attributes = True
observable_distance = normalizedDistance.getDistanceBetweenSamples(factual_sample, observable_sample, norm_type, dataset_obj)
if violating_actionable_attributes:
continue
if observable_distance < closest_observable_sample['distance']:
closest_observable_sample = {'sample': observable_sample, 'distance': observable_distance, 'norm_type': norm_type}
return closest_observable_sample
def genExp(model_trained, factual_sample, norm_type, dataset_obj):
start_time = time.time()
# SIMPLE HACK!!
# ActionSet() construction demands that all variables have a range to them. In
# the case of one-hot ordinal variables (e.g., x2_ord_0, x2_ord_1, x2_ord_2)),
# the first sub-category (i.e., x2_ord_0) will always have range(1,1), failing
# the requirements of ActionSet(). Therefore, we add a DUMMY ROW to the data-
# frame, which is a copy of another row (so not to throw off the range of other
# attributes), but setting a 0 value to all _ord_ variables. (might as well do
# this for all _cat_ variables as well).
tmp_df = dataset_obj.data_frame_kurz
sample_row = tmp_df.loc[0].to_dict()
for attr_name_kurz in dataset_obj.getOneHotAttributesNames('kurz'):
sample_row[attr_name_kurz] = 0
tmp_df = tmp_df.append(pd.Series(sample_row), ignore_index=True)
df = tmp_df
X = df.loc[:, df.columns != 'y']
# Enforce binary, categorical (including ordinal) variables only take on 2 values
custom_bounds = {
attr_name_kurz: (0, 100, 'p')
for attr_name_kurz in np.union1d(
dataset_obj.getOneHotAttributesNames('kurz'),
dataset_obj.getBinaryAttributeNames('kurz'))
}
action_set = ActionSet(X=X, custom_bounds=custom_bounds)
# action_set['x1'].mutable = False # x1 = 'Race'
# In the current implementation, we only supports any/none actionability
for attr_name_kurz in dataset_obj.getInputAttributeNames('kurz'):
attr_obj = dataset_obj.attributes_kurz[attr_name_kurz]
if attr_obj.actionability == 'none':
action_set[attr_name_kurz].mutable = False
elif attr_obj.actionability == 'any':
continue # do nothing
else:
raise ValueError(
f'Actionable Recourse does not support actionability type {attr_obj.actionability}'
)
# Enforce search over integer-based grid for integer-based variables
for attr_name_kurz in np.union1d(
dataset_obj.getIntegerBasedAttributeNames('kurz'),
dataset_obj.getBinaryAttributeNames('kurz'),
):
action_set[attr_name_kurz].step_type = "absolute"
action_set[attr_name_kurz].step_size = 1
coefficients = model_trained.coef_[0]
intercept = model_trained.intercept_[0]
if norm_type == 'one_norm':
mip_cost_type = 'total'
elif norm_type == 'infty_norm':
mip_cost_type = 'max'
else:
raise ValueError(
f'Actionable Recourse does not support norm_type {norm_type}')
factual_sample_values = list(factual_sample.values())
# p = .8
rb = RecourseBuilder(
optimizer="cplex",
coefficients=coefficients,
intercept=intercept, # - (np.log(p / (1. - p))),
action_set=action_set,
x=factual_sample_values,
mip_cost_type=mip_cost_type)
output = rb.fit()
counterfactual_sample_values = np.add(factual_sample_values,
output['actions'])
counterfactual_sample = dict(
zip(factual_sample.keys(), counterfactual_sample_values))
# factual_sample['y'] = False
# counterfactual_sample['y'] = True
counterfactual_sample['y'] = not factual_sample['y']
counterfactual_plausible = True
# IMPORTANT: no need to check for integer-based / binary-based plausibility,
# because those were set above when we said step_type = absolute! Just round!
for attr_name_kurz in np.union1d(
dataset_obj.getOneHotAttributesNames('kurz'),
dataset_obj.getBinaryAttributeNames('kurz')):
try:
assert np.isclose(counterfactual_sample[attr_name_kurz],
np.round(counterfactual_sample[attr_name_kurz]))
counterfactual_sample[attr_name_kurz] = np.round(
counterfactual_sample[attr_name_kurz])
except:
distance = -1
counterfactual_plausible = False
# return counterfactual_sample, distance
# Perform plausibility-data-type check! Remember, all ordinal variables
# have already been converted to categorical variables. It is important now
# to check that 1 (and only 1) sub-category is activated in the resulting
# counterfactual sample.
already_considered = []
for attr_name_kurz in dataset_obj.getOneHotAttributesNames('kurz'):
if attr_name_kurz not in already_considered:
siblings_kurz = dataset_obj.getSiblingsFor(attr_name_kurz)
activations_for_category = [
counterfactual_sample[attr_name_kurz]
for attr_name_kurz in siblings_kurz
]
sum_activations_for_category = np.sum(activations_for_category)
if 'cat' in dataset_obj.attributes_kurz[attr_name_kurz].attr_type:
if sum_activations_for_category == 1:
continue
else:
# print('not plausible, fixing..', end='')
# TODO: don't actually return early! Instead see the actual distance,
# fingers crossed that we can say that not only is their method giving
# counterfactuals are larger distances, but in a lot of cases they are
# not data-type plausible
# INSTEAD, do below:
# Convert to correct categorical/ordinal activations so we can
# compute the distance using already written function.
# Turns out we need to do nothing, because the distance between
# [0,1,0] and anything other than itself, (e.g., [1,1,0] or [1,0,1])
# is always 1 :)
# continue
distance = -1
counterfactual_plausible = False
# return counterfactual_sample, distance
elif 'ord' in dataset_obj.attributes_kurz[
attr_name_kurz].attr_type:
# TODO: assert activations are in order...
# if not, repeat as above...
for idx in range(int(sum_activations_for_category)):
if activations_for_category[idx] != 1:
# Convert to correct categorical/ordinal activations so we can
# compute the distance using already written function.
# Find the max index of 1 in the array, and set everything before that to 1
# print('not plausible, fixing..', end='')
# max_index_of_1 = np.where(np.array(activations_for_category) == 1)[0][-1]
# for idx2 in range(max_index_of_1 + 1):
# counterfactual_sample[siblings_kurz[idx2]] = 1
# break
distance = -1
counterfactual_plausible = False
# return counterfactual_sample, distance
else:
raise Exception(
f'{attr_name_kurz} must include either `cat` or `ord`.')
already_considered.extend(siblings_kurz)
# TODO: convert to correct categorical/ordinal activations so we can compute
# distance = output['cost'] # TODO: this must change / be verified!???
distance = normalizedDistance.getDistanceBetweenSamples(
factual_sample, counterfactual_sample, norm_type, dataset_obj)
# # TODO: post-feasibible needed???? NO
# # make plausible by rounding all non-numeric-real attributes to
# # nearest value in range
# for idx, elem in enumerate(es_instance):
# attr_name_kurz = dataset_obj.getInputAttributeNames('kurz')[idx]
# attr_obj = dataset_obj.attributes_kurz[attr_name_kurz]
# if attr_obj.attr_type != 'numeric-real':
# # round() might give a value that is NOT in plausible.
# # instead find the nearest plausible value
# es_instance[idx] = min(
# list(range(int(attr_obj.lower_bound), int(attr_obj.upper_bound) + 1)),
# key = lambda x : abs(x - es_instance[idx])
# )
end_time = time.time()
return {
'factual_sample': factual_sample,
'cfe_sample': counterfactual_sample,
'cfe_found': True, # TODO?
'cfe_plausible': counterfactual_plausible,
'cfe_distance': distance,
'cfe_time': end_time - start_time,
}
def genExp(model_trained, factual_sample, class_labels, epsilon, norm_type,
dataset_obj, standard_deviations, perform_while_plausibility):
"""
This function return the active feature tweaking vector.
x: feature vector
class_labels: list containing the all class labels
counterfactual_label: the label which we want to transform the label of x to
"""
""" initialize """
start_time = time.time()
x = np.array(list(factual_sample.values()))
# factual_sample['y'] = False
counterfactual_label = not factual_sample['y']
# initialize output in case no solution is found
closest_counterfactual_sample = dict(
zip(factual_sample.keys(), [-1 for elem in factual_sample.values()]))
closest_counterfactual_sample['y'] = counterfactual_label
counterfactual_found = False
closest_distance = 1000 # initialize cost (if no solution is found, this is returned)
# We want to support forest and tree, and we keep in mind that a trained
# tree does NOT perform the same as a trained forest with 1 tree!
if isinstance(model_trained, DecisionTreeClassifier):
# ensemble_classifier will in fact not be an ensemble, but only be a tree
estimator = model_trained
if estimator.predict(x.reshape(1, -1) != counterfactual_label):
paths_info = search_path(estimator, class_labels,
counterfactual_label)
for key in paths_info:
""" generate epsilon-satisfactory instance """
path_info = paths_info[key]
es_instance = esatisfactory_instance(x, epsilon, path_info,
standard_deviations)
if perform_while_plausibility:
# make plausible by rounding all non-numeric-real attributes to
# nearest value in range
for idx, elem in enumerate(es_instance):
attr_name_kurz = dataset_obj.getInputAttributeNames(
'kurz')[idx]
attr_obj = dataset_obj.attributes_kurz[attr_name_kurz]
if attr_obj.attr_type != 'numeric-real':
# round() might give a value that is NOT in plausible.
# instead find the nearest plausible value
es_instance[idx] = min(
list(
range(int(attr_obj.lower_bound),
int(attr_obj.upper_bound) + 1)),
key=lambda x: abs(x - es_instance[idx]))
if estimator.predict(es_instance.reshape(
1, -1)) == counterfactual_label:
counterfactual_sample = dict(
zip(factual_sample.keys(), es_instance))
counterfactual_sample['y'] = counterfactual_label
distance = normalizedDistance.getDistanceBetweenSamples(
factual_sample, counterfactual_sample, norm_type,
dataset_obj)
if distance < closest_distance:
closest_counterfactual_sample = counterfactual_sample
counterfactual_found = True
closest_distance = distance
elif isinstance(model_trained, RandomForestClassifier):
ensemble_classifier = model_trained
for estimator in ensemble_classifier:
if (ensemble_classifier.predict(x.reshape(
1, -1)) == estimator.predict(x.reshape(1, -1))
and estimator.predict(
x.reshape(1, -1) != counterfactual_label)):
paths_info = search_path(estimator, class_labels,
counterfactual_label)
for key in paths_info:
""" generate epsilon-satisfactory instance """
path_info = paths_info[key]
es_instance = esatisfactory_instance(
x, epsilon, path_info, standard_deviations)
if perform_while_plausibility:
# make plausible by rounding all non-numeric-real attributes to
# nearest value in range
for idx, elem in enumerate(es_instance):
attr_name_kurz = dataset_obj.getInputAttributeNames(
'kurz')[idx]
attr_obj = dataset_obj.attributes_kurz[
attr_name_kurz]
if attr_obj.attr_type != 'numeric-real':
# round() might give a value that is NOT in plausible.
# instead find the nearest plausible value
es_instance[idx] = min(
list(
range(int(attr_obj.lower_bound),
int(attr_obj.upper_bound) + 1)),
key=lambda x: abs(x - es_instance[idx]))
if ensemble_classifier.predict(es_instance.reshape(
1, -1)) == counterfactual_label:
counterfactual_sample = dict(
zip(factual_sample.keys(), es_instance))
counterfactual_sample['y'] = counterfactual_label
distance = normalizedDistance.getDistanceBetweenSamples(
factual_sample, counterfactual_sample, norm_type,
dataset_obj)
if distance < closest_distance:
closest_counterfactual_sample = counterfactual_sample
counterfactual_found = True
closest_distance = distance
# better naming
counterfactual_sample = closest_counterfactual_sample
distance = closest_distance
# Perform plausibility check on the nearest counterfactual found
counterfactual_plausible = True
for attr_name_kurz in dataset_obj.getInputOutputAttributeNames('kurz'):
attr_obj = dataset_obj.attributes_kurz[attr_name_kurz]
if attr_obj.attr_type != 'numeric-real':
try:
assert np.isclose(
counterfactual_sample[attr_name_kurz],
np.round(counterfactual_sample[attr_name_kurz])
), 'not satisfying plausibility (data-type)'
counterfactual_sample[attr_name_kurz] = np.round(
counterfactual_sample[attr_name_kurz])
assert counterfactual_sample[
attr_name_kurz] >= attr_obj.lower_bound, 'not satisfying plausibility (data-range)'
assert counterfactual_sample[
attr_name_kurz] <= attr_obj.upper_bound, 'not satisfying plausibility (data-range)'
except:
counterfactual_plausible = False
# distance = 1000
# return factual_sample, counterfactual_sample, distance
end_time = time.time()
return {
'factual_sample': factual_sample,
'cfe_sample': counterfactual_sample,
'cfe_found': counterfactual_found,
'cfe_plausible': counterfactual_plausible,
'cfe_distance': distance,
'cfe_time': end_time - start_time,
}
def findClosestCounterfactualSample(model_trained, model_symbols, dataset_obj,
factual_sample, norm_type, approach_string,
epsilon, log_file):
def getCenterNormThresholdInRange(lower_bound, upper_bound):
return (lower_bound + upper_bound) / 2
def assertPrediction(dict_sample, model_trained, dataset_obj):
vectorized_sample = []
for attr_name_kurz in dataset_obj.getInputAttributeNames('kurz'):
vectorized_sample.append(dict_sample[attr_name_kurz])
sklearn_prediction = int(model_trained.predict([vectorized_sample])[0])
pysmt_prediction = int(dict_sample['y'])
factual_prediction = int(factual_sample['y'])
if isinstance(model_trained, LogisticRegression):
if np.dot(model_trained.coef_,
vectorized_sample) + model_trained.intercept_ < 1e-10:
return
assert sklearn_prediction == pysmt_prediction, 'Pysmt prediction does not match sklearn prediction.'
assert sklearn_prediction != factual_prediction, 'Counterfactual and factual samples have the same prediction.'
# Convert to pysmt_sample so factual symbols can be used in formulae
factual_pysmt_sample = getPySMTSampleFromDictSample(
factual_sample, dataset_obj)
norm_lower_bound = 0
norm_upper_bound = 1
curr_norm_threshold = getCenterNormThresholdInRange(
norm_lower_bound, norm_upper_bound)
# Get and merge all constraints
print(
'Constructing initial formulas: model, counterfactual, distance, plausibility, diversity\t\t',
end='',
file=log_file)
model_formula = getModelFormula(model_symbols, model_trained)
counterfactual_formula = getCounterfactualFormula(model_symbols,
factual_pysmt_sample)
plausibility_formula = getPlausibilityFormula(model_symbols, dataset_obj,
factual_pysmt_sample,
approach_string)
distance_formula = getDistanceFormula(model_symbols, dataset_obj,
factual_pysmt_sample, norm_type,
approach_string, curr_norm_threshold)
diversity_formula = TRUE(
) # simply initialize and modify later as new counterfactuals come in
print('done.', file=log_file)
iters = 1
max_iters = 100
counterfactuals = [] # list of tuples (samples, distances)
# In case no counterfactuals are found (this could happen for a variety of
# reasons, perhaps due to non-plausibility), return a template counterfactual
counterfactuals.append({
'counterfactual_sample': {},
'counterfactual_distance': np.infty,
'interventional_sample': {},
'interventional_distance': np.infty,
'time': np.infty,
'norm_type': norm_type
})
print(
'Solving (not searching) for closest counterfactual using various distance thresholds...',
file=log_file)
while iters < max_iters and norm_upper_bound - norm_lower_bound >= epsilon:
print(
f'\tIteration #{iters:03d}: testing norm threshold {curr_norm_threshold:.6f} in range [{norm_lower_bound:.6f}, {norm_upper_bound:.6f}]...\t',
end='',
file=log_file)
iters = iters + 1
formula = And( # works for both initial iteration and all subsequent iterations
model_formula,
counterfactual_formula,
plausibility_formula,
distance_formula,
diversity_formula,
)
solver_name = "z3"
with Solver(name=solver_name) as solver:
solver.add_assertion(formula)
iteration_start_time = time.time()
solved = solver.solve()
iteration_end_time = time.time()
if solved: # joint formula is satisfiable
model = solver.get_model()
print('solution exists & found.', file=log_file)
counterfactual_pysmt_sample = {}
interventional_pysmt_sample = {}
for (symbol_key, symbol_value) in model:
# symbol_key may be 'x#', {'p0#', 'p1#'}, 'w#', or 'y'
tmp = str(symbol_key)
if 'counterfactual' in str(symbol_key):
tmp = tmp[:-15]
if tmp in dataset_obj.getInputOutputAttributeNames(
'kurz'):
counterfactual_pysmt_sample[tmp] = symbol_value
elif 'interventional' in str(symbol_key):
tmp = tmp[:-15]
if tmp in dataset_obj.getInputOutputAttributeNames(
'kurz'):
interventional_pysmt_sample[tmp] = symbol_value
elif tmp in dataset_obj.getInputOutputAttributeNames(
'kurz'): # for y variable
counterfactual_pysmt_sample[tmp] = symbol_value
interventional_pysmt_sample[tmp] = symbol_value
# Convert back from pysmt_sample to dict_sample to compute distance and save
counterfactual_sample = getDictSampleFromPySMTSample(
counterfactual_pysmt_sample, dataset_obj)
interventional_sample = getDictSampleFromPySMTSample(
interventional_pysmt_sample, dataset_obj)
# Assert samples have correct prediction label according to sklearn model
assertPrediction(counterfactual_sample, model_trained,
dataset_obj)
# of course, there is no need to assertPrediction on the interventional_sample
counterfactual_distance = normalizedDistance.getDistanceBetweenSamples(
factual_sample, counterfactual_sample, norm_type,
dataset_obj)
interventional_distance = normalizedDistance.getDistanceBetweenSamples(
factual_sample, interventional_sample, norm_type,
dataset_obj)
counterfactual_time = iteration_end_time - iteration_start_time
counterfactuals.append({
'counterfactual_sample': counterfactual_sample,
'counterfactual_distance': counterfactual_distance,
'interventional_sample': interventional_sample,
'interventional_distance': interventional_distance,
'time': counterfactual_time,
'norm_type': norm_type
})
# Update diversity and distance formulas now that we have found a solution
# TODO: I think the line below should be removed, because in successive
# reductions of delta, we should be able to re-use previous CFs
# diversity_formula = And(diversity_formula, getDiversityFormulaUpdate(model))
# IMPORTANT: something odd happens somtimes if use vanilla binary search;
# On the first iteration, with [0, 1] bounds, we may see a CF at
# d = 0.22. When we update the bounds to [0, 0.5] bounds, we
# sometimes surprisingly see a new CF at distance 0.24. We optimize
# the binary search to solve this.
norm_lower_bound = norm_lower_bound
# norm_upper_bound = curr_norm_threshold
if 'mace' in approach_string:
norm_upper_bound = float(counterfactual_distance +
epsilon / 100) # not float64
elif 'mint' in approach_string:
norm_upper_bound = float(interventional_distance +
epsilon / 100) # not float64
curr_norm_threshold = getCenterNormThresholdInRange(
norm_lower_bound, norm_upper_bound)
distance_formula = getDistanceFormula(
model_symbols, dataset_obj, factual_pysmt_sample,
norm_type, approach_string, curr_norm_threshold)
else: # no solution found in the assigned norm range --> update range and try again
with Solver(name=solver_name) as neg_solver:
neg_formula = Not(formula)
neg_solver.add_assertion(neg_formula)
neg_solved = neg_solver.solve()
if neg_solved:
print('no solution exists.', file=log_file)
norm_lower_bound = curr_norm_threshold
norm_upper_bound = norm_upper_bound
curr_norm_threshold = getCenterNormThresholdInRange(
norm_lower_bound, norm_upper_bound)
distance_formula = getDistanceFormula(
model_symbols, dataset_obj, factual_pysmt_sample,
norm_type, approach_string, curr_norm_threshold)
else:
print('no solution found (SMT issue).', file=log_file)
quit()
break
# IMPORTANT: there may be many more at this same distance! OR NONE! (what?? 2020.02.19)
closest_counterfactual_sample = sorted(
counterfactuals, key=lambda x: x['counterfactual_distance'])[0]
closest_interventional_sample = sorted(
counterfactuals, key=lambda x: x['interventional_distance'])[0]
return counterfactuals, closest_counterfactual_sample, closest_interventional_sample