MOSIC: Model-Agnostic Optimal Subgroup Identification with Multi-Constraint for Improved Reliability
Researchers have developed MOSIC, a novel framework for identifying optimal subgroups in data, particularly for applications like clinical decision-making. Unlike previous two-step methods, MOSIC employs a unified optimization approach that directly incorporates essential constraints such as subgroup size and propensity overlap. This model-agnostic method reformulates the problem into a differentiable min-max objective, solvable via gradient descent-ascent, ensuring direct constraint satisfaction during optimization. AI