PSI - Issue 78

Nicola Di Battista et al. / Procedia Structural Integrity 78 (2026) 412–417

416

Ongoing work will explore (i) refined cost-band definitions to curb class skew, (ii) synthetic minority over-sampling and cost-sensitive loss functions, and (iii) regression formulations that predict the monetary value directly, avoiding discretisation altogether. These steps are expected to enhance predictive fidelity for the most expensive repair cases while maintaining overall robustness.

4. Conclusions

This study investigated the reconstruction grant mechanism implemented in the aftermath of the 2009 L’Aquila earthquake and evaluated the feasibility of predicting reconstruction costs for masonry buildings using machine learning (ML) techniques. The predictive analysis focused on estimating the baseline grant amount allocated to dam aged buildings, expressed in € / m 2 , based on structural vulnerability indicators. The model inputs were derived from two key post-earthquake data sources: the AeDES usability classification system, which assigns buildings to one of five damage categories, and a simplified vulnerability index computed as the sum of scores attributed to nine masonry-specific structural features. These variables were extracted from a dataset comprising 2 230 masonry buildings located in eleven municipalities o ffi cially designated as the “Crater” zone—the area most severely a ff ected by the seismic event. Two classification models were trained and tested using a standard 80 / 20 hold-out procedure: a single Decision Tree (DT) and an ensemble Bagged Tree (BT) model. While the BT model outperformed the DT in most cost categories, it still showed notable weaknesses in predicting the highest cost band (1 800–2 400 € / m 2 ). This issue stems primarily from the strong class imbalance within the dataset, amplified by the uniformly spaced cost intervals used to discretize the target variable. In addition, the categorical and potentially redundant nature of the input features may have further limited model performance, particularly in capturing tail behaviours. From these preliminary results, several key lessons emerged. First, the severe imbalance across cost classes—especially in the upper grant range—negatively impacts recall, leading to underrepresentation of high-cost cases in model predictions. Second, although ensemble learning techniques e ff ectively reduce overfitting and variance, they are not su ffi cient to overcome issues related to skewed data distribution and input redundancy on their own. Future work will focus on several directions to enhance prediction accuracy and reliability. Alternative cost-class definitions, such as quantile-based banding or direct regression models, will be explored to address the imbalance problem. In parallel, cost-sensitive learning techniques and synthetic oversampling methods could be employed to improve classification of underrepresented categories. Moreover, dimensionality-reduction strategies, such as Prin cipal Component Analysis (PCA) or auto-encoders, may help uncover latent structures in the data while reducing noise. Finally, incorporating the AeDES usability classification directly as a model feature may provide additional discriminatory power for cost prediction. Ultimately, the goal is to develop a robust, data-driven ML tool that can assist public authorities in forecasting post earthquake reconstruction funding needs and enable more e ff ective planning and resource allocation at the regional scale. [1] H. Agha Beigi, T.J. Sullivan, C. Christopoulos, and G.M. Calvi. Factors influencing the repair costs of soft-story rc frame buildings and implications for their seismic retrofit. Engineering Structures , 101:233–245, 2015. cited By 28. [2] David Alexander. An evaluation of medium-term recovery processes after the 6 april 2009 earthquake in l’aquila, central italy. Environmental Hazards , 12(1):60–73, 2013. [3] Angelo Aloisio, Yuri De Santis, Francesco Irti, Dag Pasquale Pasca, Leonardo Scimia, and Massimo Fragiacomo. Machine learning predictions of code-based seismic vulnerability for reinforced concrete and masonry buildings: Insights from a 300-building database. Engineering Structures , 301:117295, 2024. [4] Angelo Aloisio, Marco Martino Rosso, Andrea Matteo De Leo, Massimo Fragiacomo, and Maria Basi. Damage classification after the 2009 l’aquila earthquake using multinomial logistic regression and neural networks. International Journal of Disaster Risk Reduction , 96:103959, 2023. [5] Angelo Aloisio, Marco Martino Rosso, Luca Di Battista, and Giuseppe Quaranta. Machine-learning-aided regional post-seismic usability prediction of buildings: 2016–2017 central italy earthquakes. Journal of Building Engineering , 91:109526, 2024. References

[6] Jack W Baker. An introduction to probabilistic seismic hazard analysis (psha). White paper, version , 1(3), 2008. [7] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning , volume 4. Springer, 2006.