The main idea is to:

• train linear models using some linear estimator M to predict each feature x_i using the protected class attribute s as input.
• then compute the residuals epsilon_ij between the predicted feature values and true feature values for each observation i for each feature j.
• The final model is then trained using epsilon_ij as input features to predict the target y.

Create a function german_dataset that outputs a pandas dataframe with the german credit data found here:

https://archive.ics.uci.edu/ml/datasets/Statlog+(German+Credit+Data)

fixes #31. This diff adds a function to read in census income data from 1994-1995, taken from https://archive.ics.uci.edu/ml/datasets/Census-Income+(KDD)

fixes #31. This diff adds a function to read in census income data from 1994-1995, taken from https://archive.ics.uci.edu/ml/datasets/Census-Income+(KDD)

fixes #31. This diff adds a function to read in census income data from 1994-1995, taken from https://archive.ics.uci.edu/ml/datasets/Census-Income+(KDD)

this revision adds constraints to the mean difference scores such that both mean difference and normalized mean difference confidence internval bounds are within -1 and 1

this revision adds constraints to the mean difference scores such that both mean difference and normalized mean difference confidence internval bounds are within -1 and 1

this revision does the following:

• add conda build recipe in the conda.recipe directory
• adds pypi install dependencies
• makefile convenience targets for building conda on 2.7 and 3.6
• adds support for python 3.6
• add travis ci for 3.6

Single Classifier Setting

• training an initial classifier on dataset D
• generating predicted probabilities on the test set
• computing the proximity of each prediction to the decision boundary learned by the classifier
• within the critical region threshold theta around the decision boundary, where 0.5 < theta < 1, X_s1 (disadvantaged observations) are assigned as y+ and X_s0 (advantaged observations are assigned as y –.

Multi-classifier Setting

ROC in the multiple classifier setting is similar to the single classifier setting, except that predicted probabilities are defined as the weighted average of probabilities generated by each classifier C_k (k is the number of different classifiers trained), where the weights can be defined as:

• the accuracy of the classifier on the data.
• uniform (take the mean of the predictions)

This technique essentially relabels the target variables using a function that can compute a decision boundary in input data space.

• the top n -ve labelled observations in the disadvantaged group s1 that are closest to the decision boundary are "promoted" to the +ve label.

• the top n +ve labelled observations in the advantaged group s0 closest to the decision boundary are "demoted' to the -ve label.

• n is the number of promotions/demotions needed to make p(+|s0) = p(+|s1)

The purpose of this issue is to add support for stratified mean difference and normalized mean difference so that we can control for other explanatory (or confounding) factors that may be driving the mean difference in outcome y between the advantaged and disadvantaged groups s_+ and s_-

DAEC is like #10, with a similar relabelling rule as ROC but re-assigns any prediction where classifiers disagree on the predicted label.

For example, if the an observation was positively labelled and the ensemble classifiers disagree on the predicted label, then the prediction would be negative.

PRR as an optimization technique that extends the standard L1/L2-norm regularization method by adding a prejudice index term to the objective function. This term is equivalent to normalized mutual information, which measures the degree to which predictions ŷ and s are dependent on each other.

With values ranging from 0 to 1, 0 means that ŷ and s are independent and a value of 1 means that they are dependent. The goal of the objective function is to find model parameters that minimize the difference between ŷ and y in addition to the degree to which ŷ depends on s. See reference below for exact implementation details.

Reference:
Kamishima, T., Akaho, S., Asoh, H., & Sakuma, J. (2012). Fairness-aware classifier with prejudice remover regularizer. Machine Learning and Knowledge Discovery in Databases, 35-50.
http://www.kamishima.net/archive/2011-ws-icdm_padm.pdf