MOOCs Dropout Prediction Based On Deep Learning

With the continuous development of information technology,more learners have chosen to acquire knowledge from the Internet.Some colleges and universities have also begun to place quality courses on the Internet for learners to learn.Providing quality college courses,this type of learning website gradually develops into a learning community,such as MOOCs(Massive open online course)platform,on which people can learn,discuss,and exam.Universities can also certificate learners according to their learning performance on the platform.MOOCs which were proposed in 2006,have been favored by a large number of online learners.Learners can study on the MOOCs platforms through the Internet in unlimited time and place.Although MOOCs develop in a favorable momentum,they also face some problems.One of the biggest problems of the MOOCs development is the phenomenon of high course dropout rates.According to statistics,only 7%-9%of learners have completed the course on MOOCs,which means that a considerable number of students have not completed the course requirements.Therefore,it is very meaningful to give early warning to the students who are going to dropout from courses according to their learning condition.But dropout prediction is a challenging task.Researchers have done a lot of studies based on traditional machine learning classification methods.However,due to the shortcomings of feature extraction items that do not reflect the characteristics of data or the lack of capacity of classification methods,traditional machines learning methods have bottlenecks in MOOCs forecasting problems.This thesis develops in a deep learning method based on the existing research.we discuss the MOOCs dropout problem,consider the multi-dimensional analysis of experimental data,the relevance of experimental data courses,the performance of absenteeism,the trend of class absenteeism,and the reasons for course absenteeism by using data visualization and fitting methods.The feature extraction items in dropout prediction are determined deliberately and we build models based on deep neural network algorithms.This thesis proposes three different training methods:direct training method,course-split training method which is based on course grouping after the association analysis and time series forecasting method which is based on time attribute characteristics of the data set.Finally,we verify the feasibility of dropout prediction of MOOCs by experiments.The experimental data set used in this thesis is derived from the KDD Cup 2015.Rich data items and millions of data volumes have brought possibilities for course dropout prediction mining accurately.Compared with the state-of-art research,the experimental results indicate that our newly proposed method boasts the following advantages:direct training method has a better predictive performance of course dropout than many classic algorithms;as an online algorithm,time series forecasting method is more suitable for the Real-time characteristics of course dropout prediction.Therefore,the course dropout prediction method proposed in this paper has a good pertinence and resolution for the MOOCs platform,and it can be directly applied to the actual MOOCs platform.