Probabilistic generative ranking method based on multi-support vector domain description

As the volume of database grows, retrieval and ordering of information according to relevance has become an important and challenging task. Ranking problem has recently been considered and formulated as a machine learning problem. Among the various learning-to-rank methods, the ranking support vector machines (SVMs) have been widely applied in various applications because of its state-of-the-art performance. In this paper, we propose a novel ranking method based on a probabilistic generative model approach. The proposed method utilizes multi-support vector domain description (multi-SVDD) and constructs pseudo-conditional probabilities for data pairs, thus enabling the construction of an efficient posterior probability function of relevance judgment of data pairs. Results of experiments on both synthetic and real large-scale datasets show that the proposed method can efficiently learn ranking functions better than ranking SVMs.