LLM-Based Evaluation Methodology of Explanation Strategies
Ege Soyarar (Özyeğin University)
Reyhan Aydogan (TU Delft - Interactive Intelligence, Özyeğin University)
Berk Buzcu (University of Applied Sciences and Arts Western Switzerland)
Davide Calvaresi (University of Applied Sciences and Arts Western Switzerland)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
As data privacy regulations, such as the EU AI Act and EU Data Act, become increasingly stringent, processing real user data for AI models like movie recommendation systems has grown more challenging. Moreover, the human-centric data collection and evaluation of Explainable AI (XAI) systems are often costly and time-consuming; making it hard to sustain. Hence, this study adopts the Synthetic Behavior Generation (SBG) approach, leveraging large language models (LLMs) to evaluate AI explanations while ensuring compliance with regulations and providing cost-effective solutions for human feedback. To assess the quality of these explanations, we utilize three different LLMs, which are fed synthetically generated user behaviors to evaluate explanations of an AI system as if they were real users. The evaluation focuses on key criteria such as convincingness, clarity, accuracy, and the impact on decision-making, facilitating a thorough assessment of explanation effectiveness. The results indicated that LLMs can deliver structured and consistent evaluations based on the provided synthetic user behavior.