| The correctness of the machine learning algorithm depends on the training samples.However,during the labeling process,the training samples will intentionally or unintentionally introduce human biases,such as gender,race and religion,and other sensitive factors.The resulting machine learning model has discrimination and prejudice against end users.This is the fairness of machine learning algorithms.As machine learning is widely used in human life,achieving equity is particularly important.In the field of fintech,machine learning has also been deeply applied.The existing research rarely considers the unfairness caused by sensitive attributes.This paper mainly considers the gender-sensitive attribute to conduct machine learning fairness research.This article first researched and analyzed the existing fairness definitions and indicators,and selected two types of fairness evaluation indicators.Comparative analysis found that there was a certain correlation between the fairness indicators,and the two types of indicators obtained from the same sample had the same distribution.Secondly,in the empirical analysis,the original machine learning algorithm(Logistic Regression,Support Vector Machine)was used for preliminary classification,and two types of fairness indicators were used to verify that the classification results without fairness did exist unfairness.Because fairness will lose accuracy to a certain degree at the same time,in order to balance fairness and accuracy,this paper introduces a trade-off coefficient λ.Through experiments,it is found that during the process of changing the value of λ from 0 to 1,fairness is improving while classification accuracy decreases.While there are fluctuations in the process.Optimal trade-off of Fairness and accuracy can be achieved when the value of λ is 0.2 and 0.6.The innovation of this paper is to verify the correlation between different fairness indicators,at the same time apply the fairness learning concept to the actual problem data set and propose a fairness logistic regression and fairness support vector machine algorithm through preprocessing. |
