Kernel Multi Label Vector Optimization (kMLVO): A Unified Multi-Label Classification Formalism

Abstract

We here propose the kMLVO (kernel Multi-Label Vector Optimization) framework designed to handle the common case in binary classification problems, where the observations, at least in part, are not given as an explicit class label, but rather as several scores which relate to the binary classification. Rather than handling each of the scores and the labeling data as separate problems, the kMLVO framework seeks a classifier which will satisfy all the corresponding constraints simultaneously. The framework can naturally handle problems where each of the scores is related differently to the classifying problem, optimizing both the classification, the regressions and the transformations into the different scores. Results from simulations and a protein docking problem in immunology are discussed, and the suggested method is shown to outperform both the corresponding SVM and SVR.