PSI - Issue 78

Angelo Aloisio et al. / Procedia Structural Integrity 78 (2026) 1–8

5

and recall versus α t for all four models. Recall for the minority class falls steeply for logistic regression, the tree and XGBoost, often below 0.20. By contrast, the ANN keeps recall above 0.60 and accuracy above 86 % even when the classes are balanced (around α t = 0 . 20). The ANN therefore o ff ers the best compromise between caution (identifying Severe cases) and overall reliability, and it can be tuned to whatever threshold a public authority chooses.

(a)

(b)

(c)

(d)

Fig. 2. Accuracy and recall of the four classifiers on validation sets, averaged over random splits, as a function of the vulnerability threshold that separates the Moderate and Severe classes.

3.3. Model interpretation

To understand the ANN’s decisions, SHAP values were computed. They measure how much each predictor pushes a single prediction up or down, averaged over all possible feature coalitions. Figure 3 ranks the inputs by mean absolute SHAP value. The most influential factors are the concrete compressive strength, building size (volume, height, area, number of storeys) and construction year. Structural typology and the presence of pilotis follow, whereas previous damage, retrofits and basement floors play a minor part. Concrete quality thus outweighs material type, underscoring the value of on-site testing in vulnerability assessment.