Canada Research Chair in Banking and Insurance Analytics

中小型企业(SME)风险管理的一个关键问题是其评估过程繁琐、低效且容易产生偏见。该程序旨在基于多模式数据(图像、文本和传统结构化数据)设计公平透明的中小企业违约概率模型。这些模式将满足关于不公平歧视法规和模式可解释性的法律要求。它将通过探索反事实公平性，即具有已知不公平影响变量的公平模型，以及创建专注于研究模型如何学习和可视化这些影响的传播方法来做到这一点。

One key issue in risk management for Small and Medium Businesses (SME) is their evaluation processes, which are cumbersome, inefficient, and bias-prone. This program aims to design fair and transparent models for the probability of default of SME based on multimodal data (images, text, and traditional structured data). These models will satisfy legal requirements regarding unfair discrimination regulations and model explainability. It will do so by exploring counterfactual fairness, i.e. fair models with a known unfair impact variable, and also by creating propagation methods that focus on studying how the models learn what they learn and visualizing these effects.