Explainable Artificial Intelligence: Counterfactual Explanations for Risk-Based Decision-Making in Construction
Jianjiang Zhan, Weili Fang, Peter E.D. Love, Hanbin Luo
Abstract
Artificial intelligence (AI) approaches, such as deep learning models, are increasingly used to determine risks in construction. However, the black-box nature of AI models makes their inner workings difficult to understand and interpret. Deploying explainable artificial intelligence (XAI) can help explain why and how the output of AI models is generated. This article addresses the following research question: <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">How can we accurately identify the critical factors influencing tunnel-induced ground settlement and provide counterfactual explanations to support risk-based decision-making?</i> We apply an XAI approach using counterfactual explanations to help understand decision-making surrounding risks when considering control ground settlement. Our approach consists of a: 1) construction of Kernel principal components analysis-based deep neural network (DNN) model; 2) generation of counterfactual explanations; 3) analysis of risk prediction and assessment factors' importance, necessity, and sufficiency. We apply our approach to the San-yang road tunnel project in Wuhan, China. The results demonstrate that the KPCA-DNN model better predicted ground settlement based on high-dimensional input features than the baseline model (i.e., AdaBoost and RandomForest). The bubble chamber pressure→ cutter-head speed→ equipment inclination is also identified as the primary risk path. Our findings indicate that using counterfactual explanations enables transparency and trust in AI-based risk models to be acquired. Moreover, our approach can help site managers, engineers, and tunnel-boring machine operators understand how to manage better and mitigate the risk of ground settlement.