PSI - Issue 78

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

3

Four classifiers are tested:(i) multinomial logistic regression [20], (ii) classification tree [8], (iii) ANN, (iv) XGBoost [10]. Model calibration uses a hold-out split; logistic regression maximises the log-likelihood with Newton–Raphson [18]. Default hyper-parameters from scikit-learn guide the other models [25].

2. Dataset overview

The dataset contains almost 300 reinforced-concrete (RC) and masonry buildings from the Province of L’Aquila, Abruzzo (Italy). Figure 1 shows where they lie and shades the mean vulnerability index (red = high, green = low; every value is below 1).

(a)

Fig. 1. Spatial distribution of the surveyed stock with colour based on the mean seismic-vulnerability index.

Table 1 lists the 15 explanatory features. Light pink cells mark categorical variables, and light green cells mark numerical ones. For concrete strength the value shown is the lowest strength measured on cores, so it applies only to RC buildings.

3. Results

This section first reviews the regression experiments and then turns to the classification study.

3.1. Regression models

Full numerical details may be found in [1]. In short, none of the four regressors, ordinary least squares, regression tree, support-vector regression (SVR) and artificial neural network (ANN), was able to reproduce the vulnerability index with acceptable accuracy on unseen data. Even though a deeper ANN drove the training R 2 up to about 0.80, its validation score fell sharply, revealing clear over-fitting. The tree model behaved similarly: moderate mean-square errors but a pronounced drop in test R 2 . SVR produced steadier errors yet only mediocre explanatory power. These