Students' Performance Prediction Based On Dimensionality Reduction

Education is the key factor to realize long-term economic and society development. Over the past few decades, in order to seek a better education system to enhance and improve the level of education, education has been reformed and innovated constantly. As the reform and innovation of education has entered the big data era, that how to choose some useful and effective information in huge amounts of information for the students,teachers,or the decision makers is really a urgent problem that needs to be resolved by the education system nowadays. In order to solve this problem,try to use dimension reduction techniques to predict student performance.Firstly, introduce four dimension reduction techniques which are principal component analysis(PCA), linear discriminate analysis(LDA), isometric map(Isomap) and local linear embedding(LLE) systematically, and summarize the advantages and disadvantages of various algorithms or research status briefly.Secondly, do basic processing for the student performance data. According to the research purpose, the data set is divided into three kinds of schemes which call A, B, C. In fact, there are six kinds of classification model in all. Then through data visualization and establishing a classification tree of the six classification respectively, found the key factors that affect student performance.Finally, simulate the models, which is also the focus of this study, and compare with the commonly method of classification. It finds that principal component analysis and locally linear embedding have better prediction effect in the condition of high dimension while linear discriminate analysis and isometric map have better prediction effect in low dimension. And compare with common classified method can find that dimension reduction techniques have better forecast effect.The main innovations of this study are two points: one is the application ofmathematical theory to show some common sense and experience of education, which are more persuasive. The other point is when dealing with classification problem, neural network, support vector machines, decision tree, and random forests are often used. However, in this study, dimension reduction techniques are used, and have better predicted effect over the common classified methods.