ChartOptimiser: Task-driven Optimisation of Chart Designs

TL;DR

ChartOptimiser tackles the fidelity gaps of LLM-generated charts by casting chart design as a task-driven, constrained optimisation problem. It uses Bayesian optimisation over an eight-dimensional VegaLite-based design space and maximises a novel four-term perceptual objective that includes white space, colour preference, task saliency, and text legibility. The approach yields faithful, efficient, and effective bar-chart designs that align with analytical tasks and compare favorably to baselines and human designs, especially on complex tasks. The work demonstrates the method's extensibility to multi-column bars and circular charts, and discusses practical implications for accessibility, localisation, and constrained-display contexts, offering a viable alternative or complement to prompt-driven LLM design.

Abstract

Automated chart design has seen significant advancements with the emergence of Large-Language Models (LLMs), which offer a practical solution for generating charts. However, LLMs frequently introduce possibly critical design failures, such as data manipulation and confabulation. While expert users can potentially mitigate these issues through iterative prompt engineering, this process requires substantial design knowledge and significant effort, remaining a massive barrier for the general public. In this paper, we present ChartOptimiser, an automated method for generating chart designs with fidelity, efficiency, and effectiveness. Given the inter-dependencies between individual design parameters, ChartOptimiser employs Bayesian optimisation to effectively search the chart design space for a novel objective function grounded in four perceptual metrics. Our empirical evaluations in bar and pie charts demonstrate that ChartOptimiser eliminates iterative design loops, providing non-expert users with high-quality charts that outperform LLM-generated designs in chart clarity, task-solving ease, and visual aesthetics.

Figures (9)

• Figure 1: Overview of ChartOptimiser. Our method operates in a design space with eight dimensions, optimising it with respect to an objective function consisting of four perceptual metrics. Bayesian optimisation is used to sample from the design space and to maximise the objective function, thus optimising the chart design for a given visual analytical task.
• Figure 2: Associations between weighted perceptual metrics (centre) used in the objective function (right) and parameters in the design space of bar charts (left).
• Figure 3: From left to right: Sample bar charts from the ChartQA masry2022chartqa dataset and corresponding bar charts optimised using ChartOptimiser, GPT-4o, LQ2 wu2021learning, as well as VegaLite default.
• Figure 4: Mean ratings and 95% confidence intervals for visual aesthetics, chart clarity, and task-solving ease for the different methods.
• Figure 5: Mean ratings and 95% confidence intervals for visual aesthetics, chart clarity, and task-solving ease for the different methods and four task types.