kingjosephm / discriminationmitigation Goto Github PK

Motivated by the Frisch-Waugh-Lovell theorem, DiscriminationMitigator offers a simple, intuitive method to mitigate potential discrimination associated with protected class features (e.g. race, gender) in supervised machine learning algorithms. In general, it is not possible to ensure a model does not, in some sense, discriminate on protected class(es) without making strong assumptions about the data generating process.
Crucially, omitting the protected class variables from the model will not necessarily prevent the model from discriminating on the basis of protected class membership. This is because other variables might proxy for protected class membership; these variables might predict a target partly through the protected class for which they proxy.

DiscriminationMitigator ameliorates proxy discrimination by taking a pre-trained ML model that was trained using the protected class attribute(s) and averages predictions across combinations of counterfactional protected class memberships. Because protected class is included in the pre-trained model, the extent to which other variables predict protected class membership does not contribute to their prediction of the target. Averaging predictions across protected class counterfactuals ensures the actual protected class membership does not contribute to the forecast of the target.

Importantly, this method does not require advanced statistical or programming knowledge, and can be used with two dominant Python ML libraries: Tensorflow Keras and LightGBM. For example code, see Example.ipynb.

Given the target variable y, a vector of protected class attributes C (c∈C), and individual i (i = 1...N):

• For each combination of protected class feature values (v∈V):
  • Set C = v for all N
  • Generate an N x 1 vector of counterfactual predictions, ŷ_v
  • Save ŷ_v to dataframe Pred_{counterfactual}
  • Repeat until end
• Pred_{counterfactual} will be N x V dimensions
• Apply weights to Pred_{counterfactual}, yielding ŷ_adjusted

• df - Pandas DataFrame to generate adjusted predictions. Typically the test set.
• model - Pre-trained LightGBM or Tensorflow Keras (Model/Sequential) model. Model must have been trained using all protected class feature(s)
• config - JSON dictionary, see Example.ipynb. This must include keys for 'protected_class_features' (i.e. a list of all protected class feature names).
• train (optional) - Pandas DataFrame used to weight predictions according to the joint distributions of protected class features in another dataset. This may be advantageous, for example, if your test set is considerably smaller than your training set.
• weights (optional) - JSON dictionary of user-specified marginal distributions, see Example.ipynb. Users can specify the custom marginals of anywhere between one and the full number of protected class features listed in config. Marginals from either df or train (if supplied) are used for the remaining protected class features, and assuming independence, the joint distributions are calculated from these. Note, feature marginals must sum to 1. The program will also raise a warning if two or more features in df appear to be one-hot vectors, reminding the user that the marginals of adjacent one-hot vectors should reflect mutual exclusivity. For example, if a binary variable is transformed into two one-hot vectors, their marginal distributions should be polar opposites. The program, however, does not enforce this; thus users must ensure the custom marginals make sense. At present it is only possible to reweight the marginal distributions of select features, not the joint distributions of combinations of features.

A Pandas Dataframe with between 2-4 columns:

• unadj_pred - unadjusted predictions, potentially containing discriminatory effect
• unif_wts - uniform weights, i.e. a simple average across all counterfactual predictions using the joint distributions of protected class features in df.
• pop_wts (optional) - if user supplies train, this will contain adjusted predictions weighted to reflect joint distributions of protected class features in that dataset.
• cust_wts (optional) - if user supplies weights, this will report the predictions adjusted according to the specified marginal distributions per feature-value combination.

See requirements.txt