A Research Of Student Performance Prediction In Online Education Based On LSTM

Student performance prediction is an important branch of online education research.It can monitor students’learning situation in real time,urge students to adjust their learning status in time,and help students better master the course content.It has very important research value.The traditional student performance prediction method has some problems,such as the classification of prediction is not detailed enough,and can not be predicted before the beginning of the course,which can not meet the current needs.In this paper,the deep learning theory is applied to study the prediction of students’performance.Based on the Long Short-Term Memory（LSTM）model,this paper studies and improves its shortcomings.The main work of this paper is as follows:（1）The traditional binary classification prediction task only predicts whether students can pass the course,but can not distinguish students’learning level.Therefore,this paper puts forward the demand of four-class classification task,that is,students’grades are divided into four categories for prediction,so as to improve the discrimination of students’final performance.（2）The traditional student performance prediction model can only predict after the beginning of the course,which has the disadvantage of lag.Therefore,this paper combines the students’background information and students’interactive information,so that the model can evaluate the students’final performance before the beginning of the course,and solve the cold start problem.（3）In this paper,Markov chain and attention mechanism are used as the strategy scheme of weight allocation.Based on LSTM model,Marcov chain-Based Multi-layer Long Short-Term Memory（MML）model and Attention-Based Multi-layer Long Short-Term Memory（AML）model are constructed respectively.The four-class classification tasks,binary classification tasks,cold start problem and generalization problem are studied for the two proposed models and three base-line models under the same experimental setting.In this paper,the accuracy and F₁score are used as the evaluation indexes.The experimental results show that the prediction effects of AML model and MML model proposed in this paper are better than the three baseline models,and the prediction effect of AML model is better than MML model.Compared with the best LSTM model in the baseline model,the accuracy of the AML model in four-class classification tasks was improved by 1.62%,and the F₁score was improved by3.40%.The accuracy of the AML model in binary classification tasks was improved by 1.73%,and the F₁score was improved by 4.10%.When facing the cold start problem,the accuracy of the AML model in four-class classification tasks was improved by 0.91%and F₁score was improved by1.31%compared with MML model.When facing the generalization problem,the accuracy of the AML model in four-class classification tasks was improved by 0.67%and F₁score was improved by 1.59%compared with MML model,the accuracy of the AML model in binary classification tasks was improved by 1.75%and F₁score was improved by 0.46%compared with MML model.