This paper investigates the application of machine learning when training a credit decision model over real, publicly available data whilst accounting for "bias objectives". We use the term "bias objective" to describe the requirement that a trained model displays discriminatory bias against a given groups of individuals that doesn't exceed a prescribed level, where such level can be zero. This research presents an empirical study examining the tension between competing model training objectives which in all cases include one or more bias objectives. This work is motivated by the observation that the parties associated with creditworthiness models have requirements that can not certainly be fully met simultaneously. The research herein seeks to highlight the impracticality of satisfying all parties' objectives, demonstrating the need for "trade-offs" to be made. The results and conclusions presented by this paper are of particular importance for all stakeholders within the credit scoring industry that rely upon artificial intelligence (AI) models as part of the decision-making process when determining the creditworthiness of individuals. This paper provides an exposition of the difficulty of training AI models that are able to simultaneously satisfy multiple bias objectives whilst maintaining acceptable levels of accuracy. Stakeholders should be aware of this difficulty and should acknowledge that some degree of discriminatory bias, across a number of protected characteristics and formulations of bias, cannot be avoided.
翻译:本文调查了在将信贷决定模式培训成真实的、公开的数据,而不是真实的、公开的数据时采用机器学习方法的情况。我们使用“偏见目标”一词来说明要求经过培训的模式对某一群体的个人表现出歧视偏见,而这种歧视并不超过规定水平,这种程度可能为零。本论文介绍了一项经验性研究,审查了在所有情况中都包含一种或多种偏见目标的相互竞争的模式培训目标之间的紧张关系。这项工作的动机是,认为与信誉模式有关的当事方有一定不能同时完全满足的要求。本研究报告试图强调满足所有当事方目标是不切实际的,表明必须作出“交易”。本论文提出的结果和结论对于信用评分行业中依赖人工智能模型作为决策过程的一部分的所有利益攸关方来说特别重要。本文件说明了培训AI模式的困难,这些模型既能同时满足多种偏见目标,同时又保持可接受的准确性。利益攸关方应认识到这一困难,并应承认某种程度的歧视性偏见,不能跨越受保护的特点和公式,避免某种程度的偏见。