Abstract: Learning to rank task is a learning process which aims at obtaining a ranking model through machine learning techniques for ranking objects. It has become one of the hot research topics in both information retrieval and machine learning communities recently. In information retrieval and machine learning fields, most of existing learning to rank approaches assume that all objects in a given query are independently and identically distributed. Although this assumption simplifies the ranking problems, the implicit interconnections among objects for each query are not exploited in the learning process. Actually, the information of the implicit interconnections can help improve the ranking performance of the ranking algorithms. In this paper, new methods are proposed in supervised ranking and transductive ranking problems to utilize the latent interconnections. In supervised ranking, a graph based ranking framework is proposed, which takes advantage of global consistency that similar objects deserve similar scores. In transductive ranking, a new query similarity measure by the interconnections among the objects is proposed, such that the more representative objects are, the more importance weighting are obtained for them. Finally, this paper validates the usefulness of relational information among objects by improving the performances of RankSVM-primal algorithm and transductive ranking algorithm in the experiments.