Doğal Afetlerde Çevrimiçi Davranış Verilerine Dayalı İhtiyaç Tahmini ve Kaynak Tahsis Optimizasyonu: 2023 Kahramanmaraş Depremleri Uygulaması

Abstract

Natural disasters are events that deeply shake the physical, economic and social structure of societies with their sudden occurrence and devastating effects. After disasters such as earthquakes, floods and forest fires, there is an intense demand for basic needs such as search and rescue, health services, food and shelter in a short time. Meeting these demands in a timely and effective manner is one of the most critical elements of the disaster management process. After a disaster, it is necessary to deliver basic aid such as search and rescue, equipment, law enforcement, health services to the regions in need. However, determining the levels of need in the regions affected by the disaster is very difficult due to limited access to information, damaged communication infrastructure and rapid change in demand dynamics. Especially if the disaster is not limited to a single center but has affected a wide geographical area, the problem of allocating resources to which regions and to what extent becomes more complex. It is not possible to wait for the determination of real demands to mobilize resources. Therefore, early and accurate prediction of potential need regions and estimated demand levels in these regions is of critical importance for the effective and timely allocation of resources. In this study, we address the demand uncertainty in natural disasters that require simultaneous intervention and the resource allocation problems that become complicated due to this uncertainty. For the solution, we develop a two-stage “predict-then-optimize” approach based on online behavioral data. We apply our approach to real data belonging to the earthquake centered in Kahramanmaraş that occurred on February 6, 2023 in the southeast of Turkey. In the first stage, the “predict” stage, we address the demand uncertainty problem using machine learning models. In this context, we first develop a demand index that uses social media shares (X, formerly known as Twitter, posts) and search engine trends (Google search trend data) in the early post-disaster period. In order to identify the data containing calls for help in X posts, we manually label 5000 X posts as “contains calls for help” (1) or “does not contain” (0). We divide the dataset into training, test and validation; We develop a text classification model based on BERT based on this structure. First, we clean the text data from punctuation marks, special characters, and user labels, and standardize the texts by converting them to lowercase. After these pre-processing steps, we compare the performance using both traditional machine learning algorithms (Support Vector Machines, Logistic Regression, Random Forest) and a BERT-based deep learning model capable of learning contextual meaning. Then, we combine X posts and Google search trend data to create a regional demand popularity index. Then, we develop Ridge regression models that determine the relative demand levels in disaster-affected regions using the index we developed and various demographic and structural variables belonging to disaster regions. While developing the regression models, we apply logarithmic, square root, inverse, square and cube transformations and binary/triple interactions to the demand popularity index and various demographic and physical variables of the provinces affected by the disaster, creating a total of 113 features. Then, in order to reduce the multicollinearity within this large feature set, we conduct a target-oriented variable elimination process with a correlation threshold of 0,95 and reduce the number of features to 33. In the modeling phase, we perform demand forecasts for February 07, 08, 09 and 10 using the Ridge regression method with multiple target variables. In the second phase, in the “optimize” phase, we develop a network flow optimization model that allocates multiple resources to disaster areas in an effective and balanced manner based on these forecasts. Thus, we determine the temporal and regional distribution plans of a limited number of teams such as search and rescue, health and logistics. In recent years, machine learning and artificial intelligence-based forecasting models have been used in decision support processes in uncertain environments such as disasters. On the other hand, resource management or allocation problems are among the basic problem classes with extensive application in the operations research literature. This thesis develops a solution approach in which prediction and optimization processes work in an integrated manner by adopting the predict-then-optimize paradigm, and applies this method using real data from the Kahramanmaraş earthquake, thus contributing to literature and practice.