Accompanied with the development of storage and processing capacity of modern technology, educational data increases sharply. It is difficult for educational researchers to derive useful information from much educational data. Therefore, educational data mining techniques are important for the development of modern education field. Recently, researches have demonstrated that machine learning, as an important tool for data mining, has shown promising performance in educational applications, especially in student performance prediction. However, few studies comprehensively compare existing machine learning methods in educational data. Moreover, most current studies only focus on a single type of educational data for student performance prediction. In this paper, three different types of task-oriented educational data are employed to investigate the performance of machine learning methods in different application scenarios. Specifically, seven parameter-optimized machine learning methods are implemented to study multiple types of performance prediction, including binary and multi-classification prediction tasks. In the experimental section, four evaluation metrics and visualizations are presented for a comparative study of different methods on three tasks, and an elaborated discussion of the experimental results is provided. The experimental results demonstrate that Random Forest has achieved superior generality on all selected datasets. In addition, the performance of Decision Tree and Artificial Neural Network models on the selected datasets indicates that they are also potential candidates to solve student performance prediction tasks.