Sparse Ranking Model Adaptation for Cross Domain Learning to Rank

Abstract

Cross domain learning to rank problem has become a hot topic issue in transfer learning and learning to rank communities. In this problem, there are scarce humanlabeled data in the target domain, but sufficient labeled data exist in a related domain named as source domain. In order to obtain an effective ranking model for target domain, various ranking adaptation frameworks are proposed to learn ranking models with the help of the labeled data in the source domain. In this paper, wepropose a sparse model adaptation framework, which utilizes ℓ1 regularization to transfer the most confident prior knowledge from source domain to the target domain. Due to the sparsity-inducing property of the ℓ1 regularization, the framework is able to reduce the negative effects of the feature gap between source domain data and target domain data. However, the optimization problem formulated by the framework is non-differentiable. It is difficult to obtain the solution by the most popular methods. To address this problem, we design an efficient algorithm from the primal-dual perspective. Empirical evaluation over LETOR benchmark data [26] collections validates that the proposed algorithm can significantly improve the accuracy of the ranking prediction.