When
Where
Assistant Professor of Information Technology Management, Georgia Institute of Technology
All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanations with Large Language Models
Abstract: Explanation of a machine learning model is an important component for gaining consumer trust in model outcomes. However, according to theoretical analysis, no single explanation method will be able to provide full, valid explanations for a given dataset and learning task. To make things worse, high-quality explanations are typically hard to obtain, and they are usually disconnected from the goal of predictions in existing explainable machine learning models. We tackle these challenges in this paper by proposing the RashomonLLM method with a series of novel design principles, where we (1) advocate the use of a collection of explanations, rather than a single explanation; (2) construct explanations using the power function class of LLMs; (3) develop a novel LLM Agentic framework of “Explanation-Prediction-Reflection” for automatic alignment between explanations and predictions. These benefits of RashomonLLM are theoretically and empirically validated in this paper, as it significantly outperforms state-of-the-art prediction models and interpretable machine learning models in two benchmark datasets. To that end, our model works well with a wide range of tabular datasets, and enables practitioners to enjoy the benefits of improving business performance and consumer trust at the same time. With our proposed model and findings, we overcome the obstacle of the accuracy-explainability tradeoff in prior methods, and we hope that practitioners will further embrace the use of XAI as a result.
Bio: Pan Li is an assistant Professor in Information Technology Management at Scheller College of Business, Georgia Tech. He was previously a visiting researcher at Google DeepMind, and he obtained his Ph.D. at the Stern School of Business, New York University. His research focuses on developing personalization techniques and XAI methods to improve consumer experiences. The results of his work were published in 18 journals and conference papers with over 1,000 citations, and some of his proposed methods were implemented by the leading tech companies, including Alibaba and Baidu. He has won the Best Dissertation award and the Best Student Paper Runner-Up award at WITS, the INFORMS Design Science Award, and the Best Reviewer Award for ISR and INFORMS Data Science Workshop.