Data-driven segmentation of observation-level logistic regression models

Abstract This study proposes a data-adaptive method to segment individual observation-based logistic regression models, focusing on motivating binary landslide data. Our method assigns observation-specific regression models and utilizes a grouped fused lasso penalty for data-adaptive model fusion when common regression coefficients are desired. However, when inherent differences persist, the models remain separate, resulting in distinct regression coefficients. To handle the large number of parameters arising from individual observation-based models, we develop a novel alternating direction method of multipliers-based algorithm. Our numerical study demonstrates improved prediction performance over conventional logistic regression models by leveraging heterogeneous data characteristics.