Eye of the Beholder: Towards Measuring Visualization Complexity.

Area of Science:

• Information Visualization
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Designing effective visualizations requires understanding perceived complexity.
• Existing research on graphical features influencing visual complexity is limited.
• Quantifying perceived visualization complexity is crucial for design evaluation.

Purpose of the Study:

• To collect human ratings of perceived visualization complexity.
• To evaluate automated methods for estimating perceived visualization complexity.
• To compare image analysis, manual feature coding, and large language model (LLM) approaches.

Main Methods:

• Conducted a crowdsourced study to gather human complexity ratings for diverse visualizations.
• Assessed image analysis metrics for correlation with human ratings.
• Developed predictive models using multilinear regression with manual feature coding.
• Evaluated a zero-shot large language model (LLM) for complexity estimation and feature extraction.

Main Results:

• Image complexity metrics showed no significant correlation with human-perceived complexity.
• Manual feature coding yielded a predictive model but was labor-intensive.
• A zero-shot LLM (GPT-4o mini) demonstrated high accuracy in rating complexity and extracting relevant features.

Conclusions:

• Perceived visualization complexity is subjective ('in the eye of the beholder').
• Zero-shot LLM prompting offers a scalable and effective approach to approximate perceived visualization complexity.
• LLMs present a promising tool for automated visualization design evaluation.