iFair: Learning Individually Fair Data Representations for Algorithmic Decision Making

subject ： Learn individual fair data representation for algorithmic decision making

This paper introduces a method of mapping user record probability to low rank representation , This method coordinates the fairness of individuals and the utility of classifiers and rankings in downstream applications . Learning low rank data representation by defining a combined loss function , This work solves the basic trade-off between utility and fairness . The author applies their method to the classification of various real data sets and learns the ranking task , Proved the universality of their method . Their experiments show that , Compared with the best preliminary work of this setting , There are substantial improvements .（ low-rank ： The rank of the matrix is low ,eg:r=1）

Research background ： In general, group equity is considered , Then the trade-off between personal fairness and accuracy has become a problem to be solved . here , A loss function is proposed to solve .

Research methods ： Map an individual to an input space , This input space is independent of a specific protected subgroup . Define mapping functions , The sum of fairness loss and accuracy loss is expressed as the objective function . This research method is called iFair frame .

Personal fair representation and calculation ： When there are two user records with all attributes xi and xj When the difference between the distance between and two records without sensitive attributes is less than or equal to the threshold , It means that there is personal fairness between these two user records .

The transformation mapping is expressed as ： among ,U Represents the previously recorded probability distribution ,V Represents the prototype vector .

The loss function includes data loss and fairness loss ：

Objective function ：

experiment ：5 Data sets

The author randomly divided the data set into three parts . Use a part of the training model to learn model parameters , The second part is the verification set , Select the superparameter by performing a grid search ( See below for details ), The third part is the test set . The same data segmentation is used in this paper to compare all methods , Altogether 6 Methods .

The evaluation index is personal fairness (yNN) And accuracy （ACC）

Experiments show that , In general iFair The effect of is better than other methods .

summary ： Our method contains two important criteria . First , We look at fairness from the perspective of application independence , This allows us to incorporate it into various tasks , Including general classifiers and regression for learning to rank . secondly , We regard personal equity as an attribute of the dataset ( In a sense , Like Privacy ), This can be achieved by preprocessing the data into a converted representation . There is no need to access protected properties at this stage . If necessary , We can also post process the representation of learning , And implement the standards of group fairness , For example, statistical equivalence .

We apply this model to five real data sets , Prove that utility and personal fairness can be coordinated to a great extent . Apply classifier and regression model to iFair In the middle , The resulting algorithm decision is more consistent than the decision made on the original data . Our method is the first method to calculate the fair result separately in the learning sorting task . For classification tasks , Its performance is better than the existing advanced work .