Assessing Bias in Risk Operations: Mitigation Strategies Through Micro Reviews

Abstract

Risk management in organizations often relies on both machine learning models and human-in-the-loop operations to detect fraudsters and identify risky users. While machine learning models excel at handling extreme cases on both ends of the risk spectrum, increasing model recall often leads to higher false positive rates. In these scenarios, automated actions are insufficient, and human reviewers play a crucial role in assessing flagged cases. However, human decisions are inherently subjective and susceptible to bias, which can impact the accuracy and fairness of risk mitigation strategies. This paper explores methods to measure the effectiveness of these human reviews, examines the types of biases that can arise during the decision-making process, and discusses strategies to mitigate these biases using MicroReviews. By integrating high-recall models with human-inthe-loop processes, we aim to develop a more balanced and unbiased approach to risk operations that enhances decision-making accuracy while minimizing potential biases.