Reducing position-sensitive subset ranking to classification

摘要: A widespread idea to attack ranking works by reducing it into a set of binary preferences and applying well studied classification techniques. The basic question addressed in this paper relates whether an accurate classifier would transfer directly good ranker. In particular, we explore reduction for subset ranking, which is based on optimization DCG metric (Discounted Cumulated Gain), standard position-sensitive performance measure. We propose consistent framework, guaranteeing that the minimal regret achievable learning pairwise assigned with importance weights. This fact allows us further develop novel upper bound terms regrets. Empirical studies benchmark datasets validate proposed approach improved performance.