A predictive assessment of households' risk against disasters caused by cold waves using machine learning

This paper trains a household-level disaster risk classifier based on supervised machine learning algorithms for cold wave-related disasters. The households' features considered for this task proxy multiple dimensions of vulnerability to disasters accounting for economic, health, social, and geographical conditions. These features are theoretically hypothesized to explain disaster risk classification. We test our predictive model based on the case of Puno, Peru, where cold wave-related disasters (e.g., −28°in 2003 and -35° in 2004) are recurrent and overwhelming. Two supervised learning algorithms were tested to build the classifiers: Logistic Regression and Random Forest Classifier. Hyperparameters of such models were optimized through Bayesian Optimization heuristic. Random Forest Classifier outperformed Logistic Regression by 1.16 % in MCC and 2.34 % in Sensitivity. In the test dataset, Random Forest Classifier achieved an MCC of 48.64 % and a Sensitivity of 80.9 %. After statistical analysis and threshold tunning, both MCC and Sensitivity increased to 51.05 % and 86.52 % in the optimal setting, respectively. Feature importance drawn from features' contribution to a reduction in entropy in the construction of the forest suggests that per capita expenditure, household localization in a rural area, altitude, and access to public goods such as paths and concrete walls drive the disaster risk classification. The model can be used to identify the demand for humanitarian aid and build operational plans to aid the most affected by cold wave-related disasters.