Traditional drug discovery methods are costly and time-consuming. The use of the existing approved drugs for treating another disease is called drug repositioning, a common strategy to overcome traditional drug discovery issues. Current drug repositioning approaches are vastly applied to machine learning methods. The performance of these methods seriously depends on types of features, their representations and the training dataset. Most drug repositioning approaches integrate different features and varied representations. As they do not consider the redundancy of features, it remains a challenge in facing drug repositioning problem. Moreover, selecting an appropriate training set is very effective in the rise of machine learning method accuracy. However, in this problem, we face two obstacles to find the proper training set. In the first one, most methods employ known drug-disease pairs as positive sets and all unknown pairs as negative sets. While the number of known pairs is much less than unknown ones, it leads to machine learning performance error because of biasing to the majority group. Second, the absence of an association between a drug and disease means this association has not been approved experimentally yet and their association may be changed. To face these obstacles, we propose DRP-VEM framework, which is assessed based on different parameters: disease feature representations, drug feature representations, classification methods, and voting ensemble training approaches.