Explanation Format does not Matter; but Explanations do - An Eggsbert study on explaining Bayesian Optimization tasks

Tanmay Chakraborty, Marion Koelle, Jörg Schlötterer, Nadine Schlicker, Christian Wirth, Christin SeifertInformation Systems Frontiers 2025 · OpenReview

Abstract

Bayesian Optimization (BO) is a family of methods for finding optimal parameters when the underlying function to be optimized is unknown. BO is used, for example, for hyperparameter tuning in machine learning and as an expert support tool for tuning cyberphysical systems. For settings where humans are involved in the tuning task, methods have been developed to explain BO (Explainable Bayesian Optimization, XBO). However, there is little guidance on how to present XBO results to humans so that they can tune the system effectively and efficiently. In this paper, we investigate how the XBO explanation format affects participants’ task performance, task load, understanding and trust in XBO. We chose a task that is accessible to a wide range of participants. Specifically, we set up an egg cooking scenario with 6 parameters that participants had to adjust to achieve a perfect soft-boiled egg. We compared three different explanation presentation formats: a bar chart, a list of rules and a textual explanation in a between-subjects online study with 213 participants. Our results show that adding any format of explanation presentation increases task success, reduces the number of trials needed to achieve success, and improves perceived understanding and confidence. While explanations add more information for participants to process, we found no increase in user task load. We also found that the aforementioned results were independent of the explanation format; all formats had a similar effect. This is an interesting finding for practical applications, as it suggests that explanations can be added to BO tuning tasks without the burden of designing or selecting specific explanation formats. In the future, it would be interesting to investigate scenarios of prolonged use of the explanation formats and whether they have different effects on participants’ mental models of the underlying system.

Introduction

Bayesian Optimization (BO) is a powerful "black-box" method used to find optimal settings for complex systems, such as tuning hyperparameters in machine learning or adjusting parameters in industrial "cyber-physical" systems. However, because BO is a black box, it is often opaque to the humans who must collaborate with it.

While the emerging field of Explainable Bayesian Optimization (XBO) aims to make these processes transparent, there is almost no research on how to present these explanations to humans. Does a data scientist prefer a chart? Does a factory operator need a text description? This research investigates whether the format of an XBO explanation affects a user's ability to perform a tuning task, their trust in the system, and their overall mental effort.

Methodology: The Eggsbert Study

The researchers conducted a between-subjects online study with 213 participants using a gamified "egg cooking" simulator. Participants had to adjust six parameters (like egg mass, initial temperature, and altitude) to achieve a perfect soft-boiled egg.

The 'Eggsbert' task interface showing AI recommendations, instructions, and parameter sliders.

The study utilized a synthetic mathematical model of egg cooking incorporating physical variables like mass, thermal properties, and the effect of altitude on boiling points. Seven unique scenarios (using different "egg types" like Ostrich or Emu) were created to ensure participants relied on the AI's guidance rather than memorization.

The researchers compared four conditions:

  1. No Explanation (Baseline): Only the AI's recommended values.
  2. Visual: A bar chart showing parameter importance and "tune/no-tune" categories.
  3. Rules: A binarized list of "If-Then" rules specifying which ranges to tune.
  4. Language: Natural language explanations generated by GPT-4 based on the underlying rules.
Comparison of the three explanation formats: (b) Visual bar chart, (c) Rule-based list, and (d) Natural language text.

To generate the core explanation data, they used TNTRules, a recent XBO algorithm that extracts interpretable patterns from the BO search space. Metrics included task success, number of trials, trust (via the Trust in Automation scale), and cognitive load (via the NASA-TLX index).

Results

The results indicate that explanations are a "free lunch" for optimization tasks. Adding explanations significantly improved task success and reduced the number of trials needed, without increasing the users' perceived task load.

  • Success and Efficiency: Participants with explanations succeeded in 54% of tasks, compared to only 29% in the no-explanation baseline. They achieved success 1.85 times more often and required significantly fewer trials.
  • Cognitive Load: NASA-TLX scores showed a significant reduction in mental demand and effort for the explanation groups, despite the additional information provided.
  • Format Agnosticism: Kruskal-Wallis tests confirmed no significant differences between the visual, rule-based, and text-based formats across all primary metrics (p > .05). This suggests that for parameter tuning, the content (the "what" and "where" to tune) is far more important than the presentation format.
  • Trust and Understanding: Users in the explanation groups reported significantly higher trust and a better subjective understanding of the system compared to the baseline.

Conclusion and Future Work

This study demonstrates that XBO tools can be integrated into existing workflows using various UI formats without sacrificing effectiveness. However, several areas remain for future exploration:

  • Complexity Scaling: Future work should examine if specific formats (like visualization) outperform others as parameter counts increase beyond the six used here.
  • Long-term Effects: Research is needed to determine if specific formats help users build better mental models over prolonged use.
  • Expert Preferences: Domain experts might have ingrained preferences for specific data formats not captured in this general audience study.

The core takeaway remains: providing truthful explanations is crucial for human-AI collaboration in optimization, while the specific format is secondary.

Summary updated

Abstract

Bayesian Optimization (BO) is a family of methods for finding optimal parameters when the underlying function to be optimized is unknown. BO is used, for example, for hyperparameter tuning in machine learning and as an expert support tool for tuning cyberphysical systems. For settings where humans are involved in the tuning task, methods have been developed to explain BO (Explainable Bayesian Optimization, XBO). However, there is little guidance on how to present XBO results to humans so that they can tune the system effectively and efficiently. In this paper, we investigate how the XBO explanation format affects participants’ task performance, task load, understanding and trust in XBO. We chose a task that is accessible to a wide range of participants. Specifically, we set up an egg cooking scenario with 6 parameters that participants had to adjust to achieve a perfect soft-boiled egg. We compared three different explanation presentation formats: a bar chart, a list of rules and a textual explanation in a between-subjects online study with 213 participants. Our results show that adding any format of explanation presentation increases task success, reduces the number of trials needed to achieve success, and improves perceived understanding and confidence. While explanations add more information for participants to process, we found no increase in user task load. We also found that the aforementioned results were independent of the explanation format; all formats had a similar effect. This is an interesting finding for practical applications, as it suggests that explanations can be added to BO tuning tasks without the burden of designing or selecting specific explanation formats. In the future, it would be interesting to investigate scenarios of prolonged use of the explanation formats and whether they have different effects on participants’ mental models of the underlying system.
BibTeX
@inproceedings{chakraborty2025explanation,
  author = {Tanmay Chakraborty and Marion Koelle and Jörg Schlötterer and Nadine Schlicker and Christian Wirth and Christin Seifert},
  title = {{Explanation Format does not Matter; but Explanations do - An Eggsbert study on explaining Bayesian Optimization tasks}},
  booktitle = {Information Systems Frontiers},
  year = {2025},
  url = {https://openreview.net/forum?id=4ptkVLbdjy},
}
Figure