Artificial Intelligence Methods for Fair and Transparent Credit Risk Rating Systems

在运筹学中，分析学——开发和使用数据驱动的人工智能技术和方法来改进组织——在支持银行业实践方面发挥着核心作用。本提案通过研究新的人工智能工具来创建公平、高效和透明的信用风险测量方法，从而在银行分析实践中向前发展，通过利用不同数据的新来源和生成集成的多模式学习人工智能模型，寻求改进目前在数据稀缺时使用的繁琐、低效和容易产生偏见的过程。虽然这些新的数据来源可能很有希望，但必须极为小心，不包括可能不公平地歧视某些人口部门的任何信息，例如包括性别或种族信息。这些信息很容易在机器学习模型中被不经意地使用，导致人们普遍担心机器学习中存在不公平的算法偏见。虽然已经提出了一些方法来创建公平的模型，但目前的技术状况并没有涵盖多模式复杂的数据源，而且在信用风险管理方面根本没有任何发展。考虑到之前的挑战，本研究将追求以下目标：-目标A：开发新的方法，以消除用于开发深度学习信用风险模型的非结构化数据（文本，图像等）中存在的潜在隐性或显性偏见。这包括性别、种族、宗教和任何其他可以在数据中识别的与身份相关的信息的影响。-目标B：构建和评估可以处理这些无偏数据的深度学习架构，并可以生成关于贷款还款概率的预测。-目标C：生成上下文相关的知识蒸馏方法，以评估在估计该概率时使用了结构化传统数据和非结构化非传统数据的哪些部分，包括图像、文本和社交网络。我将提供深度学习模型输出的可视化，并支持理解每个输入对默认估计概率的影响。对于运筹学研究学术界来说，该项目将提供新的工具来开发信用风险模型，使多模式模型可视化，并在偏差来源已知时消除偏差。在监管领域，我将提出可以采取的具体措施，以确保使用多模式数据源的公平、负责、透明和道德（FATE）信用评分系统，从而促进新的监管措施的制定。最后，在银行/金融科技界，该项目将为道德信用评分系统提供支持。

Within operational research, analytics - the development and use of data-driven artificial intelligence techniques and methodologies to improve organizations - has had a core role supporting banking practice. This proposal moves forward in banking analytics practice by researching new artificial intelligence tools to create a fair, efficient, and transparent methodology for credit risk measurement, by leveraging the new sources of diverse data and generating integrated multimodal learning artificial intelligence models, seeking to improve the cumbersome, inefficient, and bias-prone processes currently in use when data is scarce. While these new sources of data can be promising, extreme care must be taken to not include any information that would unfairly discriminate against some sectors of the population, by including e.g. gender or race information. This information can easily be inadvertently used in machine learning models, leading to the widespread concern of unfair algorithmic bias in machine learning. While some methodologies have been put forward to create fair models, the current state of the art does not cover multimodal complex data sources, and there are no developments in credit risk management at all. Considering the previous challenges, this research will pursue the following objectives: - Objective A: To develop new methodologies to remove potential implicit or explicit biases present in the unstructured data (text, images, etc) when used to develop deep learning credit risk models. This includes the effects of gender, race, religion, and any other identity-related information that can be identified in the data. - Objective B: To construct and evaluate deep learning architectures that can process this unbiased data and can generate a prediction regarding repayment probability of a loan. - Objective C: To generate context-dependent knowledge distillation approaches to evaluate what sections of the structured traditional data and unstructured non-conventional data, covering images, text, and social networks, are being used when estimating this probability. I will provide visualizations of the outputs of the deep learning models, and support the understanding of the impact of each input in the probability of default estimation. For the operational research academic community, the project will deliver new tools to develop models for credit risk, for visualizing multimodal models, and for removing biases when their sources are known. In the regulatory space, I will propose concrete measures that can be taken to ensure fair, accountable, transparent, and ethical (FATE) credit scoring systems using multimodal data sources, thus facilitating the creation of new regulatory measures. Finally, in the banking/Fintech community, the project will provide support for ethical credit scoring systems.