Predicting Student Study Failure Risk Using Artificial Neural Network Method

Background: Study failure in a course is very complicated which is always affected by many different factors and is characterized by uncertainties. We reviewed previous literatures and patents relating to student failure prediction, while there seems to be no ideal method or tool which can easily learn the relationship between failure results and reasons. Methods: In this study a method of artificial neural network was provided to predict student failure risk in course study. A three-layered network topology was used including input layer with nine neural nodes, hidden layer with ten optimized neural nodes, and output layer with one neural node, which is advantageous for dealing with this complicated and uncertain issue. The whole modeling process includes four stages: output inference, loss evaluation, weights and biases training, and model testing. Results: In our sample data there are 577 students in total, including 433 train cases and 144 test cases, and for each sample, there are nine input dependent variables and one output target variable. All calculation and optimization results were implemented based on the TensorFlow and Python. The model accuracy measurements of relative root mean square error on the total, test and train data sets were 0.1637, 0.1596, and 0.1607 respectively, and consistency was shown by testing the predicted results with our observed data, which indicated that the method was promising for predicting student failure risk in course study. Conclusion: It can be used to identify at-risk students with study difficulty and is of practical significance for educators to provide pedagogical supports and interventions in early time to the at-risk students to help them avoid academic failure, and it is also of theoretic significance to improve the whole efficiency of early warning education management.