Online Dropout Behavior Prediction Based On Multidimensional Time Series Data Analysis

MOOCs have attracted millions of learners worldwide with their flexible learning style.However,online learning is different from offline education.Lack of communication feedback leads to high dropout rates.Analyzing and predicting students’ online learning process can help teachers identify students with a tendency to drop out and provide additional help on time,which can effectively reduce the dropout rate.In this thesis two dropout prediction algorithms are investigated by analyzing the time-series data of learners’ learning behaviors on online learning platforms.1.To address the problem that the existing dropout prediction algorithms only consider the learning behavior features under a single time scale,a dropout prediction algorithm based on a multi-scale full convolutional network and a variational information bottleneck is proposed.The algorithm firstly extracts the learning behavior features of learners under different time scales using multiple parallel full convolutional networks and constructs a multi-scale full convolutional network based on this.Second,a variational information bottleneck is introduced based on the multi-scale full convolutional network to suppress the flow of irrelevant noise in the network,which reduces the influence of the noise in the temporal data of learning behaviors on the dropout prediction results.Finally,the experimental results show that the proposed algorithm achieves the best dropout prediction performance compared with the benchmark algorithm,with an F1 score and AUC values of0.922 and 0.872,respectively.Also,this thesis tests the early dropout prediction ability of the algorithm,when using learning behavior data from the first 10 days of the course,F1 scores and AUC values of 0.881 and 0.798 were obtained,respectively,and when using learning behavior data from the first 20 days of the course,F1 scores and AUC values of0.909 and 0.849 were obtained,respectively.2.An unsupervised dropout prediction algorithm based on a Gaussian mixture of variational autoencoders is proposed to address the problem that the supervised learning technique used in existing dropout prediction methods requires a large amount of manually labeled data to train the model.The algorithm first introduces a Gaussian mixture distribution in the latent variable space of the variational autoencoder so that dropouts and non-dropouts correspond to one Gaussian component in the latent variable,thus fusing feature extraction and clustering into a unified model and enabling optimization of model parameters using a backpropagation algorithm.Second,replacing the reconstruction loss of the Gaussian mixture variational autoencoder with the Soft-DTW distance improves the representation of the hidden variable for the temporal data,enabling the algorithm to extract the rich temporal features in the temporal data of learning behavior and significantly improving the dropout prediction performance of the algorithm.Experimental results show that the unsupervised dropout prediction algorithm proposed in this thesis outperforms unsupervised algorithms such as K-Means,spectral clustering,and hierarchical clustering in terms of F1 scores and ACC values,and achieves the same dropout prediction performance as classical supervised algorithms such as LDA,SVM,and CNN.This thesis empirically analyzes two dropout prediction algorithms using public datasets of online learning behaviors obtained from online learning platforms,proves the superiority and effectiveness of the proposed algorithms,and demonstrates the reliability of the experimental findings through t-tests.