Explainable AI reveals dependence modeling can outweigh aging in infrastructure resilience assessment

Dataset Description This repository contains the analytical dataset and the corresponding VAEAC-based Shapley values used in the study entitled “Explainable AI reveals dependence modeling can outweigh aging in infrastructure resilience assessment” currently under review at Nature Communications. The dataset was developed for time-dependent resilience assessment of bridges under scour-induced damage and recovery. The structured design considers six bridge ages (0, 20, 40, 60, 80, and 100 years), four environmental conditions (1: Benign, 2: Low, 3: Moderate, and 4: Severe), 12 scour-depth levels (0, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, 2.00, 2.50, 3.00, and 4.00 m), and seven span classes (1, 3, 6, 9, 12, 15, and 18 spans). These combinations define 2,016 distinct age–environment–scour–span configurations, leading to 18,432 time-dependent resilience-evaluation records per age. The damage-state variables represent the probabilities of being in no damage (DSno), minor damage (DSmin), moderate damage (DSmod), extensive damage (DSext), and severe damage (DSsev) states. The restoration variables correspond to the recovery components associated with the respective damage states Restmin, Restmod, Restext, and Restsev. These damage-state probabilities and restoration variables constitute derived features used to compute the normalized Resilience index as the target response. The Excel workbook is organized into six worksheets (Age0, Age20, Age40, Age60, Age80, and Age100) corresponding to the six considered bridge ages. Within each worksheet, columns A–M contain the 13 input variables, including time, scour depth, number of spans, environmental condition, damage-state probabilities, and restoration variables. Column N contains the target response, Resilience index. Column O contains the baseline (expected) model output used in the Shapley-value decomposition. Columns P–AB contain the VAEAC-based Shapley values corresponding, in the same order, to the 13 input variables in columns A–M. Column AC contains the Shapley-based reconstructed resilience output, obtained from the baseline value and the sum of the feature-level Shapley contributions. The Shapley values were estimated using a VAEAC [1] (Variational Autoencoder with Arbitrary Conditioning)-based conditional approach, enabling feature-attribution analysis while accounting for statistical dependence among the input variables. The dataset therefore provides, within a single workbook, both the analytical resilience data and the corresponding feature-level Shapley explanations used in the associated study. [1] Olsen, L. H., Glad, I. K., Jullum, M., & Aas, K. (2022). Using Shapley values and variational autoencoders to explain predictive models with dependent mixed features. Journal of machine learning research, 23(213), 1-51.