One vs One Mitigation of Intersectional Bias

subject ： One on one mitigation of cross deviation ： The general method of expanding fair perception binary classification

It is used to find cross deviation and reduce cross deviation .

Research background ：

Pictured 1 Shown ：

Men and women , The acceptance rate of whites and non whites is basically the same , But it is obvious that non white women have a share 0%, There is obvious discrimination , That is, a protected group as a whole seems to be treated fairly , Part of the protected group may also be treated unfairly .

Research questions ： Cross bias is the existence of two or more sensitive attributes , There is a certain correlation between these sensitive attributes , This correlation affects the results of the system , The result is unbalanced , That is, there is discrimination .

Innovation points ：1. Put forward One-vs.-One Mitigation Method

               2. A sub group parallax upper limit is provided , It can control the trade-off between accuracy and fairness .

Research methods ： In order to solve the above cross deviation problem , The paper proposes that One-vs.-One Mitigation Method .

One 、 Equity indicators

The probability that an instance in any subgroup is classified into a favorable class is equal to the probability of the entire data set        The probability that an instance in any subgroup is correctly classified into a favorable class ( The real rate ,TPR) And erroneously classified as beneficial ( false Posi- tive Rate, FPR) The probability of is equal to the probability of the entire data set

An instance in any subgroup is correctly classified as a beneficial class (TPR) The probability of is equal to the probability of instances in the entire data set

Two 、 Method

This method uses the majority voting results and prediction probability obtained from the classification model or mitigation method to calculate the score of each instance . Summarize the mitigation results of each pair , And calculate according to the average of the voting rate and prediction probability of the favorite category T. The final mitigation result Z It depends on whether the voting rate exceeds the threshold s decision

Why take a majority vote ？

Because we cannot obtain the information of prediction probability from mitigation methods such as re evaluation and different impact eliminators

The result of the majority vote means that it is with Xi Among all the results of the comparison between each pair of subgroups , The proportion of favorable classes . When there are only a few subgroups , If we only use majority voting to determine the score , Then there are many instances with the same score . This makes it difficult to determine which instances should be selected to change their classes based on the fractional value . Then we use the prediction probability extracted from the classification method to calculate the score . In order to balance majority voting results and prediction probability , We introduce a trade-off parameter w

experiment ：

MS It is massage pretreatment technology . This method is based on the prediction score of the classifier , Select the ascending and descending instances of training data ( Modify the beneficial label to the unfavorable label , vice versa ) Pre treatment .

AD It is an antagonistic depolarization , Its purpose is to minimize the possibility of predicting sensitive attribute values from prediction classes .

ROC It is a classification based on rejection of selection , As a post-treatment , Modify the class label near the classification threshold . here , The instance in the protected group is modified to a beneficial class , Instead, instances in non protected groups are modified to disadvantageous classes .

EO It is the optimal equal probability of post-processing development / Opportunity predictor , Ensure that the predicted results between groups are true 、 There is no difference in false positive rate .

Data sets ：

experimental result ：

Experimental results show that , Compared with traditional classifiers and ordinary classifiers , Our method can better reduce cross bias . We also proved that this method can control the balance between communication and fairness by adjusting the upper limit of parallax . It is proved that our method can meet a wide range of fairness requirements at the same time , Reduce the cross deviation in different real situations .

expectation ： We expect to propose new binary classification methods and standards based on fairness awareness in the future , It can reduce the burden of considering cross bias .