Optimization Challenges for Cost-Sensitive Model Prediction with Applications in Healthcare

Daniel P. Robinson, PhD, Assistant Professor, Applied Mathematics and Statistics, Whiting School of Engineering, Johns Hopkins University

I present a key challenge to incorporating user preferences into model prediction problems when the costs do not occur at the feature level of the underlying cost structure. Problems in healthcare, e.g., the prediction of adverse medical events such as Septic Shock, are important examples of such problems. During this talk I present a solution based on the definition of a group regularizer whose formulation depends on representing the underlying cost structure as a finite-layer boolean circuit. The resulting optimization problem is nonconvex and large-scale, which represents a challenge to modern optimization software. In the second half of the talk, I present a new algorithm for minimizing the sum of a convex function and a sparse regularizer, thus serves as a first step in overcoming this challenge in optimization.