Research On MOOCs Learner's Dropout Prediction Model Using Behavioral Data

Massive Open Online Courses(MOOCs)have been popular and promoted by educators all over the world.However,one of the outstanding problems in the development of MOOCs is its high dropout rate and low completion rate.In order to change this situation,the researchers have devoted their attention to the study of learners' dropout prediction.They hope that learners at dropout risk can be accurately predicted,and intervention measures can be taken in advance to make them persist in learning,so as to improve the completion rates of the courses.The dropout prediction in MOOCs has also become a hot spot in MOOCs,educational big data and educational data mining research.Although the researches on MOOCs learner's dropout prediction have made some progress,the effectiveness of the dropout prediction in practical application needs to be further improved.The key factors affecting the effectiveness of MOOCs learner's dropout prediction are:(1)the validity of data and its features,which is the premise of the effectiveness of the dropout prediction model;(2)the prediction ability of the model,which is the basic guarantee for the practicability of the dropout prediction model.Therefore,the study of this dissertation is based on the learner's behavior data which has been widely studied and proved effective in MOOCs platform,combines with the idea of ensemble learning and the related methods in deep learning,and focuses mainly on the extraction of effective features and the improvement of prediction ability of the model for MOOCs learner's dropout prediction.The main research contents and innovations of this dissertation are as follows:(1)To solve the difficulty that existing learner's dropout prediction research in MOOCs mainly relies on domain experts to extract relevant features for prediction,this dissertation proposes an integrated framework based on feature selection to predict dropout in MOOCs.The framework contains feature generation,feature selection,and dropout prediction.Specifically,a feature generation method is utilized to generate fine-grained features in days,and then an integrated feature selection method is used to select effective features that are fed into a logistic regression model for prediction.Experiments carried on a public dataset have shown that the framework can use fewer features to achieve comparable results as other dropout prediction methods in terms of precision,recall,F1 score,and AUC score,which proves the effectiveness of the extracted features.Finally,through the analysis of the effective features,some suggestions for the construction of MOOCs are proposed.(2)The extensive application of convolutional neural network proves its powerful feature extraction ability.Therefore,this dissertation attempts to use convolutional neural network to extract effective features from MOOCs learner's learning behavioral data,and propose an end-to-end dropout prediction model based on convolutional neural network.The model integrates feature extraction and classification into a framework,and improves the prediction ability of the model through their collaborative learning.The model first converts the learning behavior data with the original timestamp according to different time windows,and then uses the convolutional neural network to automatically extract the effective features from the transformed data to obtain a better classification feature representation,and finally feeds the extracted features into the classifier for classification.The non-parametric statistics of the experimental results demonstrate the effectiveness of the proposed method,which is significantly better than the existing popular methods especially in the case of large data volumes.(3)To overcome the shortcomings of the end-to-end dropout prediction model based on convolutional neural network in term of time-series features,a MOOCs learner's dropout prediction model based on neural network fusion is proposed by introducing recurrent neural network which performs well in time series data mining and combining convolutional neural network to extract local features.This model integrates the advantages of the two neural network structures,and can deal with the temporal characteristics of behavioral data more effectively than the previous model.Compared with the previous end-to-end dropout prediction model based on the convolutional neural network,the experimental results show effectiveness and superiority of the proposed model,which further expands the dropout prediction method and improves the dropout prediction effect.(4)The existing MOOCs learner's dropout prediction models mainly focus on historical data research.But there is little discussion on the prediction time for new courses,which is a problem that must be faced in practical application.In order to explore this problem,this dissertation first divides the real MOOCs learner's behavioral data set into two parts:old courses and new courses,and uses the marked historical data of old courses to train the proposed models,which is used to predict dropout with the behavioral data of new course under different time lengths.Through the analysis of the experimental results,the relationship between the behavioral data of the new courses with different time series lengths and the prediction precision,recall,F1 score,and AUC score is found,which can provide some reference for selecting an appropriate application opportunity for practical application.In summary,this dissertation mainly focuses on MOOCs learner's dropout prediction using behavioral data.To ameliorate the key stages in dropout prediction including data preprocessing,effective feature extraction and model prediction ability improvement,a series of solutions are proposed accordingly.The experimental results have proved their effectiveness and confirms that our work can provide helpful information for future MOOCs learner's dropout prediction research.