Predicting disease-related microRNAs (miRNAs) and long noncoding RNAs (lncRNAs) is crucial to find new biomarkers for the prevention, diagnosis, and treatment of complex human diseases. Computational predictions for miRNA/ lncRNA-disease associations are of great practical significance, since traditional experimental detection is expensive and timeconsuming. In this paper, we proposed a consensual machine-learning technique-based prediction approach to identify diseaserelated miRNAs and lncRNAs by high-order proximity preserved embedding (HOPE) and eXtreme Gradient Boosting (XGB), named HOPEXGB. By connecting lncRNA, miRNA, and disease nodes based on their correlations and relationships, we first created a heterogeneous disease-miRNA-lncRNA (DML) information network to achieve an effective fusion of information on similarities, correlations, and interactions among miRNAs, lncRNAs, and diseases. In addition, a more rational negative data set was generated based on the similarities of unknown associations with the known ones, so as to effectively reduce the false negative rate in the data set for model construction. By 10-fold cross-validation, HOPE shows better performance than other graph embedding methods. The final consensual HOPEXGB model yields robust performance with a mean prediction accuracy of 0.9569 and also demonstrates high sensitivity and specificity advantages compared to lncRNA/miRNA-specific predictions. Moreover, it is superior to other existing methods and gives promising performance on the external testing data, indicating that integrating the information on lncRNA-miRNA interactions and the similarities of lncRNAs/miRNAs is beneficial for improving the prediction performance of the model. Finally, case studies on lung, stomach, and breast cancers indicate that HOPEXGB could be a powerful tool for preclinical biomarker detection and bioexperiment preliminary screening for the diagnosis and prognosis of cancers. HOPEXGB is publicly available at https://github.com/airpamper/HOPEXGB.