Abstarct: Random forest, as a combination of decision trees, is one of the most comprehensive and modern methods of data mining which is placed in the category of consensus bootstrap or bagging. This method is a common tool for classification in high dimensional data, so that it is able to predict the class of observations and determine the rank of predictor variables baxsed on the importance of those variables. Random forests can be used for the quantitative, nominal or survival response variables. But when the response variable is ordinal, it is usually considered the response as a nominal one, which ignores the ordinal information and reduces the prediction accuracy. In this thesis, we use a modification of classical decision tree to construct a different type of random forest method for ordinal response variable. The results of simulation data and real data imply that by applying of these changes, the predictive performance has been improved in comparison with the classical random forest method.