Comparing verbal, visual and combined explanations for Bayesian Network inferences

TL;DR

This work addresses interpretability in Bayesian networks by introducing verbal and visual UI extensions that elucidate inference reasoning. The authors design a contribution-based explanation framework with templates, color cues, and animations to convey how findings, paths, and interactions affect target probabilities. In a controlled online study, verbal, visual, and combined explanations outperform a baseline on several question types, with the combined modality yielding the strongest gains in some cases ($p<0.001$). The findings inform BN UI design and explainable AI, and point to future avenues such as scaling to larger networks and integrating with LLM-based generation for more scalable explanations.

Abstract

Bayesian Networks (BNs) are an important tool for assisting probabilistic reasoning, but despite being considered transparent models, people have trouble understanding them. Further, current User Interfaces (UIs) still do not clarify the reasoning of BNs. To address this problem, we have designed verbal and visual extensions to the standard BN UI, which can guide users through common inference patterns. We conducted a user study to compare our verbal, visual and combined UI extensions, and a baseline UI. Our main findings are: (1) users did better with all three types of extensions than with the baseline UI for questions about the impact of an observation, the paths that enable this impact, and the way in which an observation influences the impact of other observations; and (2) using verbal and visual modalities together is better than using either modality alone for some of these question types.

Figures (19)

• Figure 1: Narrative about Podunk Beach and Podunk BN with no instantiated variables.
• Figure 2: Scale, and visual and verbal explanation for the contribution of four observations to the probability that Dermascare is present in the Podunk BN.
• Figure 3: Scale, and visual and verbal detailed explanation of the impact of the Dunkalot Sunscreen finding on the probability of Dermascare.
• Figure 4: Scale, and visual and verbal detailed explanation of the impact of the Peeling finding on the probability of Dermascare.
• Figure 5: Average mark by question type and condition.