Systems and methods for model fairness

A modelling system, comprising:

• at least one processor;
• and
• a storage device comprising machine-executable instructions that, when executed by the at least one processor, cause the modelling system to:
• obtain information identifying at least one sensitive attribute via a graphical user interface (GUI) provided to an operator device;
• execute a machine learning model building library to pre-train a predictive model on a training data set and an adversarial classifier on adversarial training data, wherein the adversarial training data include a first output generated by the predictive model and corresponding actual values for the at least one sensitive attribute;
• execute the machine learning model building library to iteratively train each of:
• an adversarial classifier that is constructed to predict a value for the at least one sensitive attribute based on a second output generated by a predictive model, wherein iteratively training the adversarial classifier comprises adjusting first parameters of the adversarial classifier to decrease a first value of a first objective function for the adversarial classifier, wherein the first objective function for the adversarial classifier is a prediction loss metric for the adversarial classifier, and
• the predictive model, wherein iteratively training the predictive model comprises adjusting second parameters of the predictive model to decrease a second value of a second objective function for the predictive model, wherein the second objective function for the predictive model is a difference between a model prediction loss metric for the predictive model and the prediction loss metric for the adversarial classifier;
• determine whether a fairness metric threshold is satisfied by the iteratively trained predictive model, wherein predictions of the iteratively trained predictive model provide for improved fairness outcomes as a result of the adversarial classifier correctly classifying one or more of the at least one sensitive attribute and thereby enabling the iteratively trained predictive model to make fairer predictions for the one or more of the at least one sensitive attribute;
• repeat the iterative training of the adversarial classifier and the predictive model using an a third output of the predictive model and an indication of at least another one of the sensitive attributes, when the determination indicates the fairness metric threshold is not satisfied, wherein the third output is generated via execution of the iteratively trained predictive model for an input data set;
• and
• provide the iteratively trained predictive model to a model execution system for deployment in a production environment, when the determination indicates the fairness metric threshold is satisfied, wherein the iteratively trained predictive model comprises a credit model configured to generate credit scores to facilitate credit decisions with respect to credit applicants.

A non-transitory computer-readable storage medium comprising machine-executable instructions that, when executed by one or more processors, cause the one or more processors to:

• execute a machine learning model building library to pre;
• train a predictive model on a training data set and an adversarial classifier on adversarial training data, wherein the adversarial training data includes a first output generated by the predictive model and corresponding actual values for at least one sensitive attribute;
• execute the machine learning model building library to iteratively train each of:
• an adversarial classifier that is constructed to predict a value for the sensitive attribute based on a second output generated by a predictive model, wherein iteratively training the adversarial classifier comprises adjusting first parameters of the adversarial classifier to decrease a first value of a first objective function for the adversarial classifier, wherein the first objective function for the adversarial classifier is a prediction loss metric for the adversarial classifier, and
• the predictive model, wherein iteratively training the predictive model comprises adjusting second parameters of the predictive model to decrease a second value of a second objective function for the predictive model, wherein the second objective function for the predictive model is a difference between a model prediction loss metric for the predictive model and the prediction loss metric for the adversarial classifier;
• determine whether a fairness metric threshold is satisfied by the iteratively trained predictive model, wherein predictions of the iteratively trained predictive model provide for improved fairness outcomes as a result of the adversarial classifier correctly classifying the sensitive attribute and thereby enabling the iteratively trained predictive model to make fairer predictions the sensitive attribute;
• repeat the iterative training of the adversarial classifier and the predictive model using a third output of the predictive model and an indication of at least another sensitive attribute, when the determination indicates the fairness metric threshold is not satisfied, wherein the third output is generated via execution of the iteratively trained predictive model for an input data set;
• and
• provide the iteratively trained predictive model to a model execution system for deployment in a production environment, when the determination indicates the fairness metric threshold is satisfied, wherein the iteratively trained predictive model comprises a credit model configured to generate credit scores to facilitate credit decisions with respect to credit applicants.