Predicting University Students' Academic Performance Based On Educational Data

Predicting student academic performance has been an important research topic in the field of Education Data Mining for a long time.Many articles have proposed predictive models using students' behavior or psychological features.However,the contents of courses also have a strong impact on students' academic performance.In this thesis,we solve the students' academic performance prediction problem from the perspective of mining the relationship between knowledge of courses.We first analyze the relationship between the basic attributes of students and courses and test scores,and further explore the correlation between the grades of courses.The results show that knowledge involved in the courses' content can affect the test scores of students.Based on this conclusion,we use students' attributes and their grades in prerequisite courses as features in the performance prediction task.The results further prove that the correlation between students' grades of the prerequisite courses can improve the accuracy of performance prediction.Based on these results,we take the knowledge of courses as an important feature to predict students' performance.We first design a knowledge extraction framework to obtain knowledge features of courses to reinforce feature groups.Comparative analyses show strong correlation between knowledge similarity and average grades of the courses in all terms,which proves that features of courses' knowledge can play a role in the prediction task.Furthermore,we propose a new model KIDNet to predict students' performance based on deep learning.The KIDNet can model higher order interactions among basic attributes of students and courses together with the knowledge features.This model uses the residual structure to accelerate the training process of neural networks and prevent over-fitting,so as to make better prediction of students' performance.Based on a real world data set,this thesis validates the performance of the proposed model in two groups of tasks,which are students' failure prediction and courses' grades prediction.The experiment results show that the courses' knowledge features play an important role in the prediction of student performance,and the proposed model KIDNet can effectively improve the accuracy of the two prediction tasks.