My Model is Unfair, Do People Even Care? Visual Design Affects Trust and Perceived Bias in Machine Learning

Area of Science:

• Human-Computer Interaction
• Machine Learning Ethics
• Data Visualization

Background:

• Machine learning (ML) models are increasingly prevalent but often contain biases.
• Stakeholders require tools to understand and evaluate ML model trade-offs, such as accuracy versus fairness.
• Visualization technology can aid in comprehending these complex model characteristics.

Purpose of the Study:

• To empirically investigate if visualization design choices influence stakeholder perception of ML model bias.
• To determine the effect of design on trust in ML models and willingness to adopt them.
• To identify user strategies for trusting ML models based on visualization.

Main Methods:

• Conducted controlled, crowd-sourced experiments with over 1,500 participants.
• Analyzed the impact of various textual and visual design choices on model perception.
• Compared how different genders prioritize fairness and performance.

Main Results:

• Gender differences in prioritizing fairness and performance were observed.
• Visualization design significantly altered fairness and performance prioritization.
• Textual explanations of fairness were more impactful than bar charts; explicit bias statements outweighed historical performance data.

Conclusions:

• Visualization design choices are critical in shaping perceptions of ML model bias and trust.
• Tailoring visualizations to user demographics and cognitive mechanisms can improve ML model adoption.
• Findings inform the design of more effective ML visualization systems for diverse stakeholders.