Exploring Commonalities in Explanation Frameworks: A Multi-Domain Survey Analysis

Eduard Barbu, Marharyta Domnich, Raul Vicente, Nikos Sakkas, André MorimarXiv 2024 · DOI · arXiv

Abstract

As machine learning models are increasingly deployed in high-stakes environments, the demand for transparency often conflicts with the pursuit of raw predictive performance. This paper presents a multi-domain survey analysis across medicine, retail, and energy sectors to identify universal requirements for Explainable AI (XAI). Our findings demonstrate a consistent willingness among stakeholders to sacrifice between 5% and 15% of model accuracy in exchange for interpretable decision-making processes. We identify counterfactual explanations and visual feature importance as the primary pillars for a cross-domain XAI framework.

Introduction

The "black box" nature of modern machine learning presents a significant barrier to adoption in regulated and high-stakes industries. While numerous XAI methods exist, they are often developed in isolation, failing to account for the diverse needs of end-users such as clinicians or retail managers. This research seeks to identify a "common denominator" for effective explanations to inform the design of universal XAI tools.

Methodology

The study employed a mixed-methods approach across three use cases:

  1. Medicine: Diabetes risk estimation and tumor progression (paraganglioma).
  2. Retail: Dynamic pricing strategies for grocery logistics.
  3. Energy: Residential energy consumption forecasting.

Researchers conducted qualitative interviews with domain experts followed by quantitative surveys using 5-point Likert scales. A technical emphasis was placed on Genetic Programming (GP), which evolves inherently readable symbolic expressions, offering a "glass-box" alternative to deep learning.

Domain-Specific Requirements

Retail Sector

In the retail domain, the focus is on operational efficiency and customer behavior. Stakeholders expressed a strong preference for simplicity. Explanations must avoid technical jargon and complex mathematical operations, favoring visual dashboards that allow for quick decision-making.

User preferences and accuracy trade-offs in the retail domain.

Medical Sector

In contrast to the retail sector, medical professionals demonstrated a higher tolerance for—and even a preference for—complexity. Clinicians valued "completeness" in explanations, trusting detailed causal models over simplified contrastive ones.

Clinician preferences for complex, causal explanations and feature distributions.

Energy Sector

Energy managers required high-level transparency to manage demand response effectively. They were the most flexible regarding performance, willing to trade up to 15% accuracy for a clear understanding of model logic. Their preferences leaned toward a combination of visual graphs and statistical summaries.

Key Results: A Universal XAI Architecture

The study proposes a dual-module architecture to satisfy cross-domain requirements:

  • Counterfactual Module: Identifies minimal input changes to flip a decision (e.g., "If energy usage dropped 10%, the bill would decrease by $X").
  • Global Importance Module: Visually ranks features (e.g., age, weather) driving the model's logic.

The Accuracy-Explainability Trade-off

A pivotal finding is the quantifiable nature of the accuracy-explainability trade-off. Users are not just asking for transparency; they are willing to pay for it in performance. The acceptable "cost" varies by domain: 5% in retail, 10% in medicine, and up to 15% in energy management.

Conclusion and Future Work

This study confirms that while domain-specific nuances exist, a universal XAI framework is feasible if centered on counterfactuals and visual importance metrics. Future research will explore the integration of Natural Language Processing (NLP) to create conversational interfaces, making "What-if" scenarios even more accessible to non-technical stakeholders.

Summary updated

BibTeX
@misc{barbu2024exploring,
  author = {Eduard Barbu and Marharyta Domnich and Raul Vicente and Nikos Sakkas and André Morim},
  title = {{Exploring Commonalities in Explanation Frameworks: A Multi-Domain Survey Analysis}},
  year = {2024},
  eprint = {2405.11958},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.11958},
}
Figure