| In recent years,with the rapid development of streaming media technology,Massive Open Online Course(MOOC)have attracted more and more peoples attention.Compared with the traditional offline teaching mode,MOOC has higher openness,and users can choose courses for learning according to their own interests and hobbies.However,due to the high degree of autonomy and selectivity and the lack of real-time interaction between teachers and students,MOOC has a very high dropout rate.The high dropout rate of MOOC causes a reduction in user scale and platform revenue,which has become one of the bottlenecks restricting the development of the MOOC platform.By analyzing and researching the log records of students’ learning behaviors,we can find out the students who may drop out in advance,and adopt manual intervention methods to them,which will help increase students’ interest in learning and course participation,reduce the dropout rate of the platform,and increase the MOOC platform income.Based on the historical behavior data of MOOC learners,this paper proposes two MOOC dropout prediction models.The main research contents are as follows:(1)Use the students learning activity records to construct data features on a weekly basis,and use the Kmeans-SMOTE algorithm to perform data over-sampling in a safe area,which can effectively avoid the generation of noisy data,thereby alleviating the unbalanced problem of dropout data samples.In addition,in order to enable the model to capture the concentration and enthusiasm of students in the course of study,this paper innovatively constructs two characteristic fields,the number of sessions and the cumulative session time.Experimental results show that the newly constructed fields can enhance the predictive ability of the model.(2)The MOOC dropout prediction is transformed into a time series prediction problem of machine learning,Based on the Attention Mechanism and Bi-directional Long-Short Term Memory(Bi-LSTM),an Att-BiLSTM dropout prediction model is proposed.By introducing an attention mechanism,the Bi-LSTM model can enhance the information capture ability of many input feature data,and focus on the feature information that has an important impact on the results of MOOC dropout.The experimental results on the real data set of KDDCup2015 show that compared with the LSTM dropout prediction model,Att-BiLSTM improves the accuracy by 0.5%.(3)Most of the existing mainstream MOOC dropout prediction models are individual dropout prediction models.Such models have the problems of low prediction accuracy and poor stability.A dropout prediction model based on stacking multiple models is proposed.First,analyze the students log activity records,and design data features on a weekly basis.Then,a two-layer integrated learning model is constructed.The first layer uses 5-fold crossvalidation to train three different individual dropout prediction models.The second layer uses Logistics Regression(LR)and combines the prediction results of the first layer.Eventually drop out predictions.Experiments on the real data set of KDDCup2015 show that the Stacking multi-model overlay dropout prediction model has better results than the individual dropout prediction model.The research in this paper shows that the dropout prediction model based on Stacking multi-model superimposed ensemble learning can reduce the risk of model overfitting and effectively enhance the stability of the dropout prediction model. |
