In the past few years, Artificial Intelligence (AI)/Machine Learning (ML) models have become prominent instruments in a multitude of sectors. However, alongside these transformative benefits, the adoption of AI/ML models has introduced challenges which could potentially increase risks in the entire model lifecycle. Thus, the inherent complexity of AI/ML models amplify the importance of comprehensive Model Risk Management (MRM) frameworks to mitigate potential risks and ensure the responsible adoption and deployment of AI/ML models in financial institutions.
Regulatory agencies across the world are also formulating frameworks to address the unique opportunities and challenges presented by Artificial Intelligence and Data Analytics (AIDA) systems. One such challenge, the existence of bias in the AI/ML models can pose several challenges that can adversely affect the decision-making process of the financial institutions using such models.
AI/ML models may result into unfair treatment in which some individuals or groups of people are privileged (i.e., receive a favourable treatment) and others are unprivileged (i.e., receive an unfavourable treatment) and decisions are based on sensitive or protected variables (such as gender, ethnicity, race, religion, disability and more). Modeling fairness in AI/ML is thus a key requirement to correct such bias in the model.
Achieving complete de-biasing of an AI/ML algorithm is simply not achievable; the objective is to reduce the presence of biases in AI/ML models. Fairness considerations should be an ongoing part of model development, model monitoring and evaluation processes.
The goal of monitoring and detecting bias is to achieve an equal probability of population groups to receive a positive treatment, or an equal treatment of individuals that only differ in sensitive/protected attributes (which partitions a population into groups whose outcomes should have parity e.g., race, religion and gender). There are some open-source toolkits available for testing fairness of the AI/ML models such as Statistical Parity, Disparate Impact and Equal opportunity difference among others.
Post bias detection, pre-processing, in-processing and post-processing bias mitigation techniques can be applied for the purpose of mitigating bias in favour of the privileged group. Post bias mitigation, it is expected that both the privileged and unprivileged group should have almost same proportion of favourable outcome, indicating bias has reduced
The empirical case study pointed out that there exists a trade-off between fairness and accuracy of the model. As the fairness is achieved in the model by reducing bias, the predictive accuracy of the model can be compromised to some extent.
While this paper provides tools for bias detection, bias mitigation and model explainability in the context of a model lifecycle, it is important to keep in mind that the notions of bias and fairness are mostly application driven or context sensitive; in other words, the choices of the attributes for measuring bias, as well as the choice of the bias metrics, can be guided by legal, social, and other non-technical considerations. The successful adoption of fairness-aware AI/ML approaches requires a thorough understanding of the characteristics of the AI/ML models in use, as well as the appropriate bias detection and mitigation algorithms. Achieving this also involves fostering collaboration across key stakeholders including AI/ML teams and end users of the models.
