Image-guided radiation treatment (IGRT) is a recent advancement in the treatment of cancer patients with tumors in the abdomen or lungs. However, the efficacy of radiation treatment in these locations is often degraded by tumor respiratory motion. Therefore, the characterization and prediction of tumor motion are critical for precise cancer radiation treatment. This paper describes an approach for knowledge discovery from respiratory motion according to different motion properties. A hierarchical data model is proposed for tumor motion data representation. Various statistical analysis and correlation discovery over complex tumor respiratory motion data are designed based on a data cube to characterize different tumor motion properties. The outcomes will provide quantitative information for tumor motion prediction and real-time treatment delivery, which results better care for cancer patients.