A data binning-based multi-task prediction method for earthquake casualty prediction

Reliable earthquake casualty estimates can help governments make quick and precise resource allocation decisions for victim rescue. Although machine learning has been widely applied to casualty prediction, large disparities across earthquake cases and limitedavailable data hinder accurate modeling. This paper proposes a data binning-based multi-task prediction method that enahnces forecasting through task decomposition and adaptive estimator optimization. A data binning tree algorithm first partitions samples into bins with minimal target dispersion by identifying shared attribute patterns. Learning strategies are then tailored by bin size:  for small bins, prediction rules are derived from the tree structure; for big bins, regression models are adaptively constructed using a customized evolutionary algorithm to select the optimal model, parameters, and feature subset. For any new case, we first identify the corresponding bin and then apply the associated estimator. Using real-world earthquake cases, comprehensive experiments verify that the proposed method outperforms several state-of-the-art methods in casualty prediction.