Detecting and mitigating human bias in visual analytics

Abstract

People are susceptible to a multitude of biases, including perceptual biases and illusions; cognitive biases like confirmation bias or anchoring bias; and social biases like racial or gender bias that are borne of cultural experiences and stereotypes. As humans are an integral part of data analysis and decision making in many domains, their biases can be injected into and even amplified by models and algorithms. This dissertation focuses on developing a better understanding of the role of human biases in visual data analysis. It is comprised of three high-level goals: 1. Define bias: We present four common perspectives on the term “bias” and describe how they are relevant in the context of visual data analysis. 2. Detect bias: We introduce a set of computational bias metrics that, applied to user interaction sequences in real-time, can be used to approximate bias in the user’s analysis process. 3. Mitigate bias: We describe a design space of ways in which visualizations might be modified to increase awareness of bias. We implement a system which integrates and visualizes the bias metrics and show how it can increase awareness of bias.