An interpretable two-stage adaptive deep learning model for humanitarian aid information prediction and emergency response support

Diverse modes of information in social media posts during emergency responses collectively present an opportunity to advance artificial intelligence (AI) technologies to promote the integration of AI in humanitarian aid operations. To accurately identify humanitarian aid information and its categories, and to facilitate effective emergency responses, we first designed a two-stage humanitarian aid information prediction framework (THAIP). The first stage identifies humanitarian aid information and the second stage predicts the specific categories of information. We then developed an interpretable two-stage adaptive deep learning model (ITADL) based on THAIP, which adaptively determines the optimal data modality, model structure, and parameters based on the tasks at different stages. When applied to a real-world dataset from the social media platform Twitter in the context of emergency response, THAIP and ITADL achieved superior performance compared to models using a single-stage framework and several other deep learning models. Furthermore, the responses predicted by ITADL are interpreted at both global and local levels, enhancing the model's interpretability and providing valuable decision support for humanitarian aid planning and emergency response.