FairUP / src /aif360 /datasets /structured_dataset.py
erasmopurif's picture
First commit
d2a8669
raw
history blame
25.7 kB
from collections import defaultdict
from contextlib import contextmanager
from copy import deepcopy
from logging import warning
import numpy as np
import pandas as pd
from aif360.datasets import Dataset
class StructuredDataset(Dataset):
"""Base class for all structured datasets.
A StructuredDataset requires data to be stored in :obj:`numpy.ndarray`
objects with :obj:`~numpy.dtype` as :obj:`~numpy.float64`.
Attributes:
features (numpy.ndarray): Dataset features for each instance.
labels (numpy.ndarray): Generic label corresponding to each instance
(could be ground-truth, predicted, cluster assignments, etc.).
scores (numpy.ndarray): Probability score associated with each label.
Same shape as `labels`. Only valid for binary labels (this includes
one-hot categorical labels as well).
protected_attributes (numpy.ndarray): A subset of `features` for which
fairness is desired.
feature_names (list(str)): Names describing each dataset feature.
label_names (list(str)): Names describing each label.
protected_attribute_names (list(str)): A subset of `feature_names`
corresponding to `protected_attributes`.
privileged_protected_attributes (list(numpy.ndarray)): A subset of
protected attribute values which are considered privileged from a
fairness perspective.
unprivileged_protected_attributes (list(numpy.ndarray)): The remaining
possible protected attribute values which are not included in
`privileged_protected_attributes`.
instance_names (list(str)): Indentifiers for each instance. Sequential
integers by default.
instance_weights (numpy.ndarray): Weighting for each instance. All
equal (ones) by default. Pursuant to standard practice in social
science data, 1 means one person or entity. These weights are hence
person or entity multipliers (see:
https://www.ibm.com/support/knowledgecenter/en/SS3RA7_15.0.0/com.ibm.spss.modeler.help/netezza_decisiontrees_weights.htm)
These weights *may not* be normalized to sum to 1 across the entire
dataset, rather the nominal (default) weight of each entity/record
in the data is 1. This is similar in spirit to the person weight in
census microdata samples.
https://www.census.gov/programs-surveys/acs/technical-documentation/pums/about.html
ignore_fields (set(str)): Attribute names to ignore when doing equality
comparisons. Always at least contains `'metadata'`.
metadata (dict): Details about the creation of this dataset. For
example::
{
'transformer': 'Dataset.__init__',
'params': kwargs,
'previous': None
}
"""
def __init__(self, df, label_names, protected_attribute_names,
instance_weights_name=None, scores_names=[],
unprivileged_protected_attributes=[],
privileged_protected_attributes=[], metadata=None):
"""
Args:
df (pandas.DataFrame): Input DataFrame with features, labels, and
protected attributes. Values should be preprocessed
to remove NAs and make all data numerical. Index values are
taken as instance names.
label_names (iterable): Names of the label columns in `df`.
protected_attribute_names (iterable): List of names corresponding to
protected attribute columns in `df`.
instance_weights_name (optional): Column name in `df` corresponding
to instance weights. If not provided, `instance_weights` will be
all set to 1.
unprivileged_protected_attributes (optional): If not provided, all
but the highest numerical value of each protected attribute will
be considered not privileged.
privileged_protected_attributes (optional): If not provided, the
highest numerical value of each protected attribute will be
considered privileged.
metadata (optional): Additional metadata to append.
Raises:
TypeError: Certain fields must be np.ndarrays as specified in the
class description.
ValueError: ndarray shapes must match.
"""
if df is None:
raise TypeError("Must provide a pandas DataFrame representing "
"the data (features, labels, protected attributes)")
#if df.isna().any().any():
# raise ValueError("Input DataFrames cannot contain NA values.")
#try:
# df = df.astype(np.float64)
#except ValueError as e:
# print("ValueError: {}".format(e))
# raise ValueError("DataFrame values must be numerical.")
# Convert all column names to strings
df.columns = df.columns.astype(str).tolist()
label_names = list(map(str, label_names))
protected_attribute_names = list(map(str, protected_attribute_names))
self.feature_names = [n for n in df.columns if n not in label_names
and (not scores_names or n not in scores_names)
and n != instance_weights_name]
self.label_names = label_names
self.features = df[self.feature_names].values.copy()
self.labels = df[self.label_names].values.copy()
self.instance_names = df.index.astype(str).tolist()
if scores_names:
self.scores = df[scores_names].values.copy()
else:
self.scores = self.labels.copy()
df_prot = df.loc[:, protected_attribute_names]
self.protected_attribute_names = df_prot.columns.astype(str).tolist()
self.protected_attributes = df_prot.values.copy()
# Infer the privileged and unprivileged values in not provided
if unprivileged_protected_attributes and privileged_protected_attributes:
self.unprivileged_protected_attributes = unprivileged_protected_attributes
self.privileged_protected_attributes = privileged_protected_attributes
else:
self.unprivileged_protected_attributes = [
np.sort(np.unique(df_prot[attr].values))[:-1]
for attr in self.protected_attribute_names]
self.privileged_protected_attributes = [
np.sort(np.unique(df_prot[attr].values))[-1:]
for attr in self.protected_attribute_names]
if instance_weights_name:
self.instance_weights = df[instance_weights_name].values.copy()
else:
self.instance_weights = np.ones_like(self.instance_names,
dtype=np.float64)
# always ignore metadata and ignore_fields
self.ignore_fields = {'metadata', 'ignore_fields'}
# sets metadata
super(StructuredDataset, self).__init__(df=df, label_names=label_names,
protected_attribute_names=protected_attribute_names,
instance_weights_name=instance_weights_name,
unprivileged_protected_attributes=unprivileged_protected_attributes,
privileged_protected_attributes=privileged_protected_attributes,
metadata=metadata)
def subset(self, indexes):
""" Subset of dataset based on position
Args:
indexes: iterable which contains row indexes
Returns:
`StructuredDataset`: subset of dataset based on indexes
"""
# convert each element of indexes to string
indexes_str = [self.instance_names[i] for i in indexes]
subset = self.copy()
subset.instance_names = indexes_str
subset.features = self.features[indexes]
subset.labels = self.labels[indexes]
subset.instance_weights = self.instance_weights[indexes]
subset.protected_attributes = self.protected_attributes[indexes]
subset.scores = self.scores[indexes]
return subset
def __eq__(self, other):
"""Equality comparison for StructuredDatasets.
Note: Compares all fields other than those specified in `ignore_fields`.
"""
if not isinstance(other, StructuredDataset):
return False
def _eq(x, y):
if isinstance(x, np.ndarray) and isinstance(y, np.ndarray):
return np.all(x == y)
elif isinstance(x, list) and isinstance(y, list):
return len(x) == len(y) and all(_eq(xi, yi) for xi, yi in zip(x, y))
return x == y
return all(_eq(self.__dict__[k], other.__dict__[k])
for k in self.__dict__.keys() if k not in self.ignore_fields)
def __ne__(self, other):
return not self == other
def __repr__(self):
# return repr(self.metadata)
return str(self)
def __str__(self):
df, _ = self.convert_to_dataframe()
df.insert(0, 'instance_weights', self.instance_weights)
highest_level = ['instance weights'] + \
['features']*len(self.feature_names) + \
['labels']*len(self.label_names)
middle_level = [''] + \
['protected attribute'
if f in self.protected_attribute_names else ''
for f in self.feature_names] + \
['']*len(self.label_names)
lowest_level = [''] + self.feature_names + ['']*len(self.label_names)
df.columns = pd.MultiIndex.from_arrays(
[highest_level, middle_level, lowest_level])
df.index.name = 'instance names'
return str(df)
# TODO: *_names checks
def validate_dataset(self):
"""Error checking and type validation.
Raises:
TypeError: Certain fields must be np.ndarrays as specified in the
class description.
ValueError: ndarray shapes must match.
"""
super(StructuredDataset, self).validate_dataset()
# =========================== TYPE CHECKING ============================
for f in [self.features, self.protected_attributes, self.labels,
self.scores, self.instance_weights]:
if not isinstance(f, np.ndarray):
raise TypeError("'{}' must be an np.ndarray.".format(f.__name__))
# convert ndarrays to float64
self.features = self.features.astype(np.float64)
self.protected_attributes = self.protected_attributes.astype(np.float64)
self.labels = self.labels.astype(np.float64)
self.instance_weights = self.instance_weights.astype(np.float64)
# =========================== SHAPE CHECKING ===========================
if len(self.labels.shape) == 1:
self.labels = self.labels.reshape((-1, 1))
try:
self.scores.reshape(self.labels.shape)
except ValueError as e:
print("ValueError: {}".format(e))
raise ValueError("'scores' should have the same shape as 'labels'.")
if not self.labels.shape[0] == self.features.shape[0]:
raise ValueError("Number of labels must match number of instances:"
"\n\tlabels.shape = {}\n\tfeatures.shape = {}".format(
self.labels.shape, self.features.shape))
if not self.instance_weights.shape[0] == self.features.shape[0]:
raise ValueError("Number of weights must match number of instances:"
"\n\tinstance_weights.shape = {}\n\tfeatures.shape = {}".format(
self.instance_weights.shape, self.features.shape))
# =========================== VALUE CHECKING ===========================
if np.any(np.logical_or(self.scores < 0., self.scores > 1.)):
warning("'scores' has no well-defined meaning out of range [0, 1].")
for i in range(len(self.privileged_protected_attributes)):
priv = set(self.privileged_protected_attributes[i])
unpriv = set(self.unprivileged_protected_attributes[i])
# check for duplicates
if priv & unpriv:
raise ValueError("'privileged_protected_attributes' and "
"'unprivileged_protected_attributes' should not share any "
"common elements:\n\tBoth contain {} for feature {}".format(
list(priv & unpriv), self.protected_attribute_names[i]))
# check for unclassified values
if not set(self.protected_attributes[:, i]) <= (priv | unpriv):
raise ValueError("All observed values for protected attributes "
"should be designated as either privileged or unprivileged:"
"\n\t{} not designated for feature {}".format(
list(set(self.protected_attributes[:, i])
- (priv | unpriv)),
self.protected_attribute_names[i]))
# warn for unobserved values
if not (priv | unpriv) <= set(self.protected_attributes[:, i]):
warning("{} listed but not observed for feature {}".format(
list((priv | unpriv) - set(self.protected_attributes[:, i])),
self.protected_attribute_names[i]))
@contextmanager
def temporarily_ignore(self, *fields):
"""Temporarily add the fields provided to `ignore_fields`.
To be used in a `with` statement. Upon completing the `with` block,
`ignore_fields` is restored to its original value.
Args:
*fields: Additional fields to ignore for equality comparison within
the scope of this context manager, e.g.
`temporarily_ignore('features', 'labels')`. The temporary
`ignore_fields` attribute is the union of the old attribute and
the set of these fields.
Examples:
>>> sd = StructuredDataset(...)
>>> modified = sd.copy()
>>> modified.labels = sd.labels + 1
>>> assert sd != modified
>>> with sd.temporarily_ignore('labels'):
>>> assert sd == modified
>>> assert 'labels' not in sd.ignore_fields
"""
old_ignore = deepcopy(self.ignore_fields)
self.ignore_fields |= set(fields)
try:
yield
finally:
self.ignore_fields = old_ignore
def align_datasets(self, other):
"""Align the other dataset features, labels and protected_attributes to
this dataset.
Args:
other (StructuredDataset): Other dataset that needs to be aligned
Returns:
StructuredDataset: New aligned dataset
"""
if (set(self.feature_names) != set(other.feature_names) or
set(self.label_names) != set(other.label_names) or
set(self.protected_attribute_names)
!= set(other.protected_attribute_names)):
raise ValueError(
"feature_names, label_names, and protected_attribute_names "
"should match between this and other dataset.")
# New dataset
new = other.copy()
# re-order the columns of the new dataset
feat_inds = [new.feature_names.index(f) for f in self.feature_names]
label_inds = [new.label_names.index(f) for f in self.label_names]
prot_inds = [new.protected_attribute_names.index(f)
for f in self.protected_attribute_names]
new.features = new.features[:, feat_inds]
new.labels = new.labels[:, label_inds]
new.scores = new.scores[:, label_inds]
new.protected_attributes = new.protected_attributes[:, prot_inds]
new.privileged_protected_attributes = [
new.privileged_protected_attributes[i] for i in prot_inds]
new.unprivileged_protected_attributes = [
new.unprivileged_protected_attributes[i] for i in prot_inds]
new.feature_names = deepcopy(self.feature_names)
new.label_names = deepcopy(self.label_names)
new.protected_attribute_names = deepcopy(self.protected_attribute_names)
return new
# TODO: Should we store the protected attributes as a separate dataframe
def convert_to_dataframe(self, de_dummy_code=False, sep='=',
set_category=True):
"""Convert the StructuredDataset to a :obj:`pandas.DataFrame`.
Args:
de_dummy_code (bool): Performs de_dummy_coding, converting dummy-
coded columns to categories. If `de_dummy_code` is `True` and
this dataset contains mappings for label and/or protected
attribute values to strings in the `metadata`, this method will
convert those as well.
sep (char): Separator between the prefix in the dummy indicators and
the dummy-coded categorical levels.
set_category (bool): Set the de-dummy coded features to categorical
type.
Returns:
(pandas.DataFrame, dict):
* `pandas.DataFrame`: Equivalent dataframe for a dataset. All
columns will have only numeric values. The
`protected_attributes` field in the dataset will override the
values in the `features` field.
* `dict`: Attributes. Will contain additional information pulled
from the dataset such as `feature_names`, `label_names`,
`protected_attribute_names`, `instance_names`,
`instance_weights`, `privileged_protected_attributes`,
`unprivileged_protected_attributes`. The metadata will not be
returned.
"""
df = pd.DataFrame(np.hstack((self.features, self.labels)),
columns=self.feature_names+self.label_names,
index=self.instance_names)
df.loc[:, self.protected_attribute_names] = self.protected_attributes
# De-dummy code if necessary
if de_dummy_code:
df = self._de_dummy_code_df(df, sep=sep, set_category=set_category)
if 'label_maps' in self.metadata:
for i, label in enumerate(self.label_names):
df[label] = df[label].replace(self.metadata['label_maps'][i])
if 'protected_attribute_maps' in self.metadata:
for i, prot_attr in enumerate(self.protected_attribute_names):
df[prot_attr] = df[prot_attr].replace(
self.metadata['protected_attribute_maps'][i])
# Attributes
attributes = {
"feature_names": self.feature_names,
"label_names": self.label_names,
"protected_attribute_names": self.protected_attribute_names,
"instance_names": self.instance_names,
"instance_weights": self.instance_weights,
"privileged_protected_attributes": self.privileged_protected_attributes,
"unprivileged_protected_attributes": self.unprivileged_protected_attributes
}
return df, attributes
def export_dataset(self, export_metadata=False):
"""
Export the dataset and supporting attributes
TODO: The preferred file format is HDF
"""
if export_metadata:
raise NotImplementedError("The option to export metadata has not been implemented yet")
return None
def import_dataset(self, import_metadata=False):
""" Import the dataset and supporting attributes
TODO: The preferred file format is HDF
"""
if import_metadata:
raise NotImplementedError("The option to import metadata has not been implemented yet")
return None
def split(self, num_or_size_splits, shuffle=False, seed=None):
"""Split this dataset into multiple partitions.
Args:
num_or_size_splits (array or int): If `num_or_size_splits` is an
int, *k*, the value is the number of equal-sized folds to make
(if *k* does not evenly divide the dataset these folds are
approximately equal-sized). If `num_or_size_splits` is an array
of type int, the values are taken as the indices at which to
split the dataset. If the values are floats (< 1.), they are
considered to be fractional proportions of the dataset at which
to split.
shuffle (bool, optional): Randomly shuffle the dataset before
splitting.
seed (int or array_like): Takes the same argument as
:func:`numpy.random.seed()`.
Returns:
list: Splits. Contains *k* or `len(num_or_size_splits) + 1`
datasets depending on `num_or_size_splits`.
"""
# Set seed
if seed is not None:
np.random.seed(seed)
n = self.features.shape[0]
if isinstance(num_or_size_splits, list):
num_folds = len(num_or_size_splits) + 1
if num_folds > 1 and all(x <= 1. for x in num_or_size_splits):
num_or_size_splits = [int(x * n) for x in num_or_size_splits]
else:
num_folds = num_or_size_splits
order = list(np.random.permutation(n) if shuffle else range(n))
folds = [self.copy() for _ in range(num_folds)]
features = np.array_split(self.features[order], num_or_size_splits)
labels = np.array_split(self.labels[order], num_or_size_splits)
scores = np.array_split(self.scores[order], num_or_size_splits)
protected_attributes = np.array_split(self.protected_attributes[order],
num_or_size_splits)
instance_weights = np.array_split(self.instance_weights[order],
num_or_size_splits)
instance_names = np.array_split(np.array(self.instance_names)[order],
num_or_size_splits)
for fold, feats, labs, scors, prot_attrs, inst_wgts, inst_name in zip(
folds, features, labels, scores, protected_attributes, instance_weights,
instance_names):
fold.features = feats
fold.labels = labs
fold.scores = scors
fold.protected_attributes = prot_attrs
fold.instance_weights = inst_wgts
fold.instance_names = list(map(str, inst_name))
fold.metadata = fold.metadata.copy()
fold.metadata.update({
'transformer': '{}.split'.format(type(self).__name__),
'params': {'num_or_size_splits': num_or_size_splits,
'shuffle': shuffle},
'previous': [self]
})
return folds
@staticmethod
def _de_dummy_code_df(df, sep="=", set_category=False):
"""De-dummy code a dummy-coded dataframe obtained with pd.get_dummies().
After reversing dummy coding the corresponding fields will be converted
to categorical.
Args:
df (pandas.DataFrame): Input dummy coded dataframe
sep (char): Separator between base name and dummy code
set_category (bool): Set the de-dummy coded features
to categorical type
Examples:
>>> columns = ["Age", "Gender=Male", "Gender=Female"]
>>> df = pd.DataFrame([[10, 1, 0], [20, 0, 1]], columns=columns)
>>> _de_dummy_code_df(df, sep="=")
Age Gender
0 10 Male
1 20 Female
"""
feature_names_dum_d, feature_names_nodum = \
StructuredDataset._parse_feature_names(df.columns)
df_new = pd.DataFrame(index=df.index,
columns=feature_names_nodum + list(feature_names_dum_d.keys()))
for fname in feature_names_nodum:
df_new[fname] = df[fname].values.copy()
for fname, vl in feature_names_dum_d.items():
for v in vl:
df_new.loc[df[fname+sep+str(v)] == 1, fname] = str(v)
if set_category:
for fname in feature_names_dum_d.keys():
df_new[fname] = df_new[fname].astype('category')
return df_new
@staticmethod
def _parse_feature_names(feature_names, sep="="):
"""Parse feature names to ordinary and dummy coded candidates.
Args:
feature_names (list): Names of features
sep (char): Separator to designate the dummy coded category in the
feature name
Returns:
(dict, list):
* feature_names_dum_d (dict): Keys are the base feature names
and values are the categories.
* feature_names_nodum (list): Non-dummy coded feature names.
Examples:
>>> feature_names = ["Age", "Gender=Male", "Gender=Female"]
>>> StructuredDataset._parse_feature_names(feature_names, sep="=")
(defaultdict(<type 'list'>, {'Gender': ['Male', 'Female']}), ['Age'])
"""
feature_names_dum_d = defaultdict(list)
feature_names_nodum = list()
for fname in feature_names:
if sep in fname:
fname_dum, v = fname.split(sep, 1)
feature_names_dum_d[fname_dum].append(v)
else:
feature_names_nodum.append(fname)
return feature_names_dum_d, feature_names_nodum