Convex covariate clustering for classification

Clustering, like covariate selection for classification, is an important step to compress and interpret the data. However, clustering of covariates often performed independently classification step, which can lead undesirable results that harm interpretability compression rate. Therefore, we propose a method cluster while taking into account class label information samples. We formulate problem as convex optimization uses both, a-priori similarity between covariates, from class-labeled Like ordinary [1], proposed offers unique global minima making it insensitive initialization. In order solve problem, specialized alternating direction multipliers (ADMM), scales up several thousands variables. Furthermore, in circumvent computationally expensive cross-validation, model criterion based on approximating marginal likelihood. Experiments synthetic real data confirm usefulness criterion.