Research On Student Performance Prediction Method Based On Hybrid Decomposition Technology

With the gradual development of educational informationalization, the education data increase rapidly. Under the network teaching circumstance, it can guide course design and teaching methods, because it could record students' learning behavior data by mining student' log files, interactive information, learner' information. It could be benefit for improving teacher'teaching ability if we could predict student performance by exiting data. Educational data mining is the process that extract meaningful information in educational settings by data mining method. This information could provide effective service to learners, teachers and managers. This paper belongs to educational data mining field, and relates to predict student performance furthermore, which is based on blended matrix factorization technology.First, based on the in-depth study about factorization technology knowledge and theory, we propose a blended matrix factorization method. The first step of this method is dividing Student-Task matrix into two smaller matrices by using matrix factorization technology. The second step, we add a regularization term in order to prevent overfitting problem. The third step, we compute local minima between predictive value and true value by stochastic gradient descent method. The fourth step is using biases metric factorization to solve student bias and task bias.Second, using tensor factorization, this study takes the temporal effect into account. From the education point of view, the knowledge of the learners will be cumulated over the time; therefore, the original two-mode Student-Task matrix can be divided to three-mode' Student-Task-Time tensor. We reduce the dimension of the tensor by CP method, and then use the moving average forecasting method to solve the impact of temporal effect.Finally, in order to test the predicting efficiency of the proposed blended matrix factorization method, this study use real world data set, which is collected from the Knowledge Discovery and Data Mining Challenge 2010, to do research and analysis. Making comparison of experimental result between our method and existing methods, this study analyzes their strengths, weaknesses and the applicable condition in detail. In the end, we make a detailed summary about the predict student performance based on blended matrix factorization method, and then make a prospect of the future work.This study can apply to educational recommender system, this system can predict learners' other tasks performance by analyzing the exiting data, and then recommend appropriate tasks for learners to improve teachers' teaching efficiency and learners' learning efficiency.