MOOC Dropout Prediction Algorithm Based On Deep Neural Network

With the popularization of Internet in various fields,the traditional education and teaching methods are changing.Since the birth of Massive Open Online Courses(MOOC)in 2012,it has attracted more and more people participating in online learning.But the high dropout rate has brought inconvenience to the management of MOOC platform teaching.Predicting learners' dropping out behavior is helpful to improve learners' learning effect and realize more value of MOOC platform.Based on the learning behavior characteristics of MOOC learners,this thesis proposes several deep network-based dropout prediction algorithms.The main research contents are as follows:(1)A dropout prediction model based on convolutional long short-term memory(CNN-LSTM)network is proposed.Aiming at solving the problems of cumbersome feature extraction process,unstable prediction effect and insufficient prediction ability of the existing MOOC dropout prediction models,this model adopts a feature automatic extraction strategy without specific domain knowledge and manual intervention,and takes into account the long-term dependence information of students' behavior characteristics to improve the prediction ability of the model.The experimental results show that,compared with the LSTM and CNN-RNN dropout prediction models,the model improves the AUC by 2.7% and 1.4%,respectively,.(2)A dropout prediction model that uses long short-term memory based on richer convolutional features(RCF-LSTM)network is proposed.Aiming at solving the problems of large time span,coarse-grained features,and insufficient utilization of CNN's rich hierarchical features,the model takes finer-grained behavioral features as input while also effectively blending the fine-grained features of each CNN layer.The experimental results show that,compared with the CNN-RNN and CNN-LSTM dropout prediction models,the AUC of this model is improved by 0.29% and 0.25% respectively.(3)A dropout prediction model based on CNN-LSTM-SVM network is proposed.Aiming at solving the problems of multiple training parameters,memory consumption and data disequilibrium,the model comprehensively takes the local and sequential characteristics of student behaviors into consideration,and sets different weights for different categories.The model solves the influence of unbalanced data on prediction results.Compared with the dropout prediction algorithm based on the RCF-LSTM network proposed in Chapter 3,this model is simple in structure and improves the AUC by 5.54%.The researches in this thesis show that the deep learning dropout prediction models can predict the behavioral changes of students at different steps.It can help teachers to provide timely interventions,improve students' learning effects,which is of great guiding significance for promoting the rapid development of education informationization of China.