深度学习模型中的公平性研究

近几年深度神经网络正被广泛应用于现实决策系统,决策系统中的不公平现象会加剧社会不平等,造成社会危害.因此研究者们开始对深度学习系统的公平性展开大量研究,但大部分研究都从群体公平的角度切入,且这些缓解群体偏见的方法无法保证群体内部的公平.针对以上问题,本文定义了两种个体公平率计算方法,分别为基于输出标签的个体公平率（IFR_b）,即相似样本对在模型预测中标签相同的概率和基于输出分布的个体公平率（IFR_p）,即相似样本对的预测分布差异在阈值范围内的概率,后者是更严格的个体公平.更进一步,本文提出一种提高模型个体公平性的算法IIFR,该算法通过余弦相似度计算样本之间的差异程度,利用相似临界值筛选出满足条件的相似训练样本对,最后在训练过程中将相似训练样本对的输出差异作为个体公平损失项添加到目标函数中,惩罚模型输出差异过大的相似训练样本对,以达到提高模型个体公平性的目的.实验结果表明,IIFR算法在个体公平的提升上优于最先进的个体公平提升方法.此外IIFR算法能够在提高模型个体公平性的同时,较好地维持模型的群体公平性.

Research on Fairness in Deep Learning Models

In recent years, deep neural networks have been widely used in real decision-making systems. Inequities in decision-making systems will exacerbate social inequality and cause social harm. Therefore, researchers have begun to carry out a lot of research on the fairness of deep learning systems, but most of them focus on group fairness, and they cannot guarantee the fairness within the group. In response to the above problems, we define two individual fairness calculation methods, which are individual fairness rate IFR_b based on labels of output, that is the probability of having same predicted label for two similar samples, and individual fairness rate IFR_p based on distributions of output, that is the probability of having similiar predicted output distribution for two similar samples, respectively, the latter being the stricter individual fairness. In addition, we also propose an algorithm IIFR to improve the individual fairness of these models. The algorithm uses the cosine similarity to measure the similarity between samples, and then selects the similar sample pairs by the similarity threshold decided by different applications, finally adds the output difference of the similar sample pairs to the objective function as an individual fairness loss item during the training process, which penalizes the similar training samples with large differences of model output in order to improve the individual fairness of the model. The experimental results show that our IIFR algorithm outperforms the state-of-the-art method on the improvement of individual fairness. In addition, IIFR can maintain group fairness of models while improving individual fairness.