Funding for training human resources in most countries is very important and costly. Hence in training, prediction of students with expelled risky is one of the today's key issues and researches. There are imbalances in the training data that causes reduce prediction accuracy in fail students. In this paper, experiments based on data mining techniques have been tried to improve prediction accuracy of fail students. To do this, data from the UCI site are used that contains 5820 records in

more »

... students. First, the training data are clustered to select the most appropriate algorithm, Farthest First, and then by doing experiments, best Features including the fitness level of instructor and students' attendance level and the best Test option, 90% for training data, are selected. Questionnaire is used to Weight features. Finally, the cost-sensitive classification algorithms have been implemented with the proposed cost matrix and the model provided the best results. Results prove that this model can play an important role in promoting science education centers with an accuracy rate of 96.47%, TP rate 99.2% and precision rate of 96%.