A reduced listwise approach of learning to rank

Ranking functions determine the relevance of search results in information retrieval systems. Recently learning to rank has become a promising method for constructing a model or a function for ranking objects. Several listwise approaches were proposed as an alternate of learning to rank algorithms, which directly define a loss function on list of objects. Motivated by demonstrating the effectiveness using ListMLE, a new reduced approach called RcList is introduced. A regularization condition is appended to RcList, and imposes constraints on the model parameter space. The unique optimal solution of the optimization object of the RcList is obtained by applying Newton-YUAN method. It is demonstrated the performance benefits of the RcList compared to the ListMLE through experiments on a public dataset LETOR.