ibm / infairness Goto Github PK

method link: https://github.com/IBM/inFairness/blob/main/inFairness/utils/datautils.py#L9

Hi,

I was trying to run the basic example provided here: https://hub.gke2.mybinder.org/user/ibm-infairness-h9o8a789/doc/workspaces/auto-6/tree/examples/adult-income-prediction/adult_income_prediction.ipynb
However I am running into issue :

TypeError Traceback (most recent call last)
Input In [10], in <cell line: 8>()
5 distance_x = distances.LogisticRegSensitiveSubspace()
6 distance_y = distances.SquaredEuclideanDistance()
----> 8 distance_x.fit(X_train, protected_idxs)
9 distance_y.fit(num_dims=output_size)
11 distance_x.to(device)

File /srv/conda/envs/notebook/lib/python3.8/site-packages/inFairness/distances/logistic_sensitive_subspace.py:77, in LogisticRegSensitiveSubspace.fit(self, data_X, data_SensitiveAttrs, protected_idxs, keep_protected_idxs, autoinfer_device)
42 """Fit Logistic Regression Sensitive Subspace distance metric
43
44 Parameters
(...)
73 on CPU.
74 """
76 if data_SensitiveAttrs is not None and protected_idxs is None:
---> 77 basis_vectors_ = self.compute_basis_vectors_data(
78 X_train=data_X, y_train=data_SensitiveAttrs
79 )
81 elif data_SensitiveAttrs is None and protected_idxs is not None:
82 basis_vectors_ = self.compute_basis_vectors_protected_idxs(
83 data_X,
84 protected_idxs=protected_idxs,
85 keep_protected_idxs=keep_protected_idxs,
86 )

File /srv/conda/envs/notebook/lib/python3.8/site-packages/inFairness/distances/logistic_sensitive_subspace.py:164, in LogisticRegSensitiveSubspace.compute_basis_vectors_data(self, X_train, y_train)
161 X_train = datautils.convert_tensor_to_numpy(X_train)
162 y_train = datautils.convert_tensor_to_numpy(y_train)
--> 164 self.assert_sensitiveattrs_binary(y_train)
166 basis_vectors_ = []
167 outdim = y_train.shape[-1]

File /srv/conda/envs/notebook/lib/python3.8/site-packages/inFairness/distances/logistic_sensitive_subspace.py:185, in LogisticRegSensitiveSubspace.assert_sensitiveattrs_binary(self, sensitive_attrs)
183 def assert_sensitiveattrs_binary(self, sensitive_attrs):
--> 185 assert validationutils.is_tensor_binary(
186 sensitive_attrs
187 ), "Sensitive attributes are required to be binary to learn the metric. Please binarize these attributes before fitting the metric."

File /srv/conda/envs/notebook/lib/python3.8/site-packages/inFairness/utils/validationutils.py:19, in is_tensor_binary(data)
5 """Checks if the data is binary (0/1) or not. Return True if it is binary data
6
7 Parameters
(...)
15 True if data is binary. False if not
16 """
18 nonbindata = (data != 0) & (data != 1)
---> 19 has_nonbin_data = True in nonbindata
20 return not has_nonbin_data

TypeError: argument of type 'bool' is not iterable

Google Colab currently provides Python 3.7 (bug) which inFairness does not support. Binder supports specifying python versions.

include default values but expose all parameters to the end users for all classes and methods.

• Logistic regression metric: model
• SVD: expose subspace

Geomloss has a build problem and right now we are pointing to a fork.

Also, the output prediction should include the rankings, not only the scores

• Currently, the logistic regression sensitive subspace metric initializes the logistic regression model in a fixed way.
• Proposed change is to allow the user to specify custom arguments to the model

paper link: https://arxiv.org/abs/2103.11023

Currently, the Logistic Regression metric requires the complete data (with protected attributes) as input along with protected indices. This should be changed to accepting the datasets without protected attributes as X and the protected attributes data as Y

https://github.com/IBM/inFairness/blob/main/inFairness/distances/logistic_sensitive_subspace.py#L20-L26

retiring adults dataset: https://github.com/zykls/folktables

Travis CI builds terminating citing: "Job has been terminated due to insufficient credits balance". Investigate migrating to GitHub actions.

paper: https://arxiv.org/abs/2110.13796