Object Feature Coding: A Decomposition Framework Unifying Object and Label Ranking

Object ranking refers to the problem of learning to order a set of objects, each object described by an attribute tuple. Label ranking refers to the problem of learning to order a set of objects, each object only described by a nominal label; the order depends on a context, which is represented by an attribute tuple. The present work seeks to unify both approaches through the means of problem decomposition. To this end, a framework is proposed which allows to split the original problem into multiple sub-problems, which are then solved and their solutions aggregated into a solution to the original problem. The decomposition of the problem is a function of the object features; many different methods of decomposition are possible within the framework, and several are presented. The introduced methods are examined and compared to established or more obvious approaches. Along the way, several properties of the different types of ranking tasks are discussed.