XAI tools in the public sector: a case study on predicting combined sewer overflows
Nicholas Maltbie, Nan Niu, Matthew Van Doren, Reese Johnson
Abstract
Artificial intelligence and deep learning are becoming increasingly prevalent in contemporary software solutions. Explainable artificial intelligence (XAI) tools attempt to address the black box nature of the deep learning models and make them more understandable to humans. In this work, we apply three state-of-the-art XAI tools in a real-world case study. Our study focuses on predicting combined sewer overflow events for a municipal wastewater treatment organization. Through a data driven inquiry, we collect both qualitative information via stakeholder interviews and quantitative measures. These help us assess the predictive accuracy of the XAI tools, as well as the simplicity, soundness, and insightfulness of the produced explanations. Our results not only show the varying degrees that the XAI tools meet the requirements, but also highlight that domain experts can draw new insights from complex explanations that may differ from their previous expectations.