Introduction to bias and fairness in ML models

Measures of fairness:

Metrics based on base rates:

• Disparate Impact(DI): ratio between the probability of unprivileged group gets favorable prediction and the probability of privileged group gets favorable prediction
• Statistical Parity Difference(SPD): similar to DI but instead of ratios, differences is calculated

Metrics based on group conditioned rates:

• Equal Opportunity Difference (EOD): difference between TPR values of unprivileged and privileged groups.
• Average Odds Difference (AOD): average of false positive rate difference between FPR of unprivileged and privileged groups and TPR of unprivileged and privileged groups.
• Error Rate Difference (ERD):
• Error rate ERR=FPR+FNR
• ERD= ERR(U)-ERR(P)

Metrics based on individual fairness:

• Consistency (CNT): measures how similar are the predictions when the instances are similar.
• Theil Index (TI) / Entropy Index: Measures both group and individual fairness.

Bias Mitigation Techniques:

Pre-Processing Algorithms: do not change the model, only works on dataset before training

• Reweighing: different weights are assigned to reduce effect of favouritism of a specified group.
• Disparate Impact Remover (DIR): based on the concept of DI. It modifies the value of protected attribute to remove distinguishing factors

In-Processing Algorithms: modify ML model

• Adversarial Debiasing: introduces backward feedback(negative gradient) for predicting protected attribute which is achieved by using adversarial model that learns from difference between protected and other attributes.
• Prejudice Remover Regularizer: if a model’s decision is dependent on a protected attribute, it is called a direct prejudice. To handle this , we can remove this protected variable or regulate its effect on ml model. This regularization is used under this approach where a regularizer is implemented that computes the effect of protected attribute.

Post-Processing Algorithms: modifies the predicted results instead of ML models or input data

• Equalized odds (E): it changes the output labels to optimize EOD metric. A linear program is solved to obtain probabilities of modifying prediction.
• Calibrated Equalized odds: this optimizes EOD metric by using calibrated prediction score produced by classifier.
• Reject Option Classification: it favors the instances in privileged group over unprivileged ones that lie in the decision boundary with high uncertainty.

Source:

Do the Machine Learning Models on a Crowd Sourced Platform Exhibit Bias? An Empirical Study on Model Fairness Sumon Biswas Dept. of Computer Science, Iowa State University Ames, IA, USA sumon@iastate.edu Hridesh Rajan Dept. of Computer Science, Iowa State University Ames, IA, USA hridesh@iastate.edu