This paper delves into the pivotal role of machine learning in responding to natural disasters and understanding human behavior during crises. Natural disasters, from earthquakes to floods, have profound consequences for both the environment and society, impacting health, the economy, and mental well-being. Prevention and preparedness are key components of disaster management, yet the psychological challenges faced by affected individuals are equally significant. Psychosocial support and educational programs play a vital role in aiding individuals in their recovery. Machine learning, in this context, offers the ability to predict the evolution of natural disasters, providing early warnings that can save lives and reduce losses. It further extends to analyzing data related to human behavior during disasters, enhancing readiness for future calamities. This study specifically addresses the challenge of understanding human behavior during a snowstorm that struck Greece in 2023, employing artificial intelligence techniques to develop classification models categorizing individuals into three distinct groups based on socio-economic characteristics and is one of the few machine learning approaches that have been performed to date on data derived from corresponding questionnaire surveys. Artificial intelligence methodologies were harnessed to construct these classification models, with a focus on categorizing individuals into three specific classes: “Did not travel at all”, “Traveled only as necessary”, or “Did not limit travel”. The dataset employed in this study was collected through a survey conducted within the framework of the AEGIS+ research project, concentrating on assessing the mental health of individuals impacted by natural disasters. The goal was to generalize the optimal classification model and extract knowledge applicable in natural disaster scenarios. Three methodological frameworks for data analysis were proposed, incorporating combinations of Simple Logistic Regression and Inductive Decision Trees with the SMOTE data balancing method and a new data balancing method called LCC (Leveling of Cases per Class), within the context of validation procedures like “Use Train Set,” “10-fold Cross Validation,” and “Hold Out.” This paper’s contribution lies in the development of hybrid classification models, highlighting the significance of data balancing with LCC method throughout the modeling process. The results were deemed satisfactory, with the inductive decision tree method demonstrating superior performance (Classification accuracy near to 90%). This approach, offering strong classification rules, holds potential for knowledge application in natural disaster risk management. Knowledge Mining and Metadata Analysis further revealed the socio-economic characteristics influencing the decision to move during a natural disaster, including age, education, work-status, and workstyle. Crucially, this work, in addition to providing knowledge through the data mining process that can be used to estimate evacuation probability, develop targeted emergency information messages, and improve evacuation planning, is also used as a catalyst for future research efforts. It encourages the collection of relevant data, the exploration of new challenges in data analysis related to natural disasters and mental health, and the development of new data balancing methods and hybrid data analysis methodological frameworks.
Understanding evacuation behavior for effective disaster preparedness: a hybrid machine learning approach
