Hybrid integration of TOPSIS and machine learning models in a scenario-aware framework for shelter prioritization under conflict conditions

Determining shelter sites in conflict-affected areas poses a major challenge for humanitarian agencies, faced with time limitations, limited data, and uncertainty. This study presents a Hybrid Scenario-Aware Decision Support Framework that integrates Multi-Criteria Decision-Making (MCDM) and Machine Learning (ML), offering transparent, scalable, and data-driven shelter evaluation. At first, the TOPSIS technique was used to rank shelter sites based on indicators of accessibility, occupancy capacity, safety from exposure to conflict, and severity of humanitarian need. These rankings were then employed as target outputs for ten of the regression-based ML models, which include the Linear regression (LR), the Ridge regression (RR), the Elastic net regression (ENR), the Squared exponential Gaussian process regression (SEGPR), and ensemble models like the Random Forest (RF) and Boosted trees (BT). They systematically assessed the framework in Essential, Rapid, and Comprehensive scenarios, with increasing complexity based on information from northern Syria. The results show that SEGPR achieved the best predictive performance in the first and second scenarios, with maximum predictive performance, as it recorded the lowest RMSE of 0.12656 in the first scenario and of 0.09371 in the second scenario, along with a decent NSE score. The third scenario, the ENR, was observed to have slightly better accuracy, indicating the dependence of the models on scenarios. Tree-based models were more destabilized under data-poverty. Overall, the proposed framework effectively combines the transparency of MCDM with the adaptive predictive capability of ML, enabling robust performance across varying operational contexts. By reducing reliance on purely subjective decision-making and enhancing scalability under crisis constraints, the developed toolkit offers a transferable, scenario-sensitive solution for humanitarian planners engaged in shelter prioritization and emergency logistics in conflict-affected environments. Unlike existing hybrid optimization-based frameworks that rely on static weighting or predefined aggregation mechanisms, the proposed approach introduces a scenario-aware predictive layer that dynamically adapts to the intensity of conflicts while minimizing subjective weighting bias through data-driven learning.