Online Ranking Learning on Clusters

摘要

Online data stream ranking learning problem is considered using training data in the form of a sequence of identical items series, described by a number of features and relative rank within the series. It is assumed that feature values and relative ranks of the same items may vary slightly for different series of observations, and there are stable groups of items with similar properties. In this regard, the problem of learning to rank on clusters is stated, while training dataset consist of estimates of centers of clusters and average rank of the items inside each cluster. A unified approach to ranking learning on clusters using kernel models of utility function is proposed. Recurrent algorithms for estimating the parameters of a utility function model as well as recurrent ranking learning algorithm in the space of conjugate variables are developed.