Research Of Students’ Academic Early Warning Based On Multi-Source Online Behavior Data

With the gradual rise of online education,education data mining aims to extract valuable information from huge amounts of data on the Internet and help educators make more humanized educational decisions.Among them,using data mining technology to predict students’ academic performances and early warnings of risk students are the most basic and important part.However,there are still some disadvantages in the related research.Firstly,the existing research of academic early warning usually obtains the relevant data of students’ learning behavior from a single data source for prediction,ignoring the impact of students’ overall online behavior pattern and external environment on students’ academic performance.Secondly,there are only a few risk students in the student group,but the risk student prediction model has little research on this kind of unbalanced data.In addition,although more and more deep learning models are used in academic early warning,they cannot reasonably explain and attribute the predictions,which is not conducive to the understanding and application of final users.This thesis mainly aims at the shortcomings of the existing research and the main work and innovation are as follows:(1)Using multi-source data to model students’ behavior,a sequential prediction based on deep network(SPDN)is proposed.The framework makes full use of two data sources of online learning platform records and campus network logs to extract students’ online learning behavior and daily network use pattern.The multi-channel convolution network structure can fuse multi-source data,extract the behavioral features automatically and combine with the static characteristics of students.The long and short-term memory network is used to model the overall online behavior of students,and finally predict the probability of risk students.Experimental results show that the performance of the proposed model is better than the baselines.(2)In order to find a very small number of students with academic risk in imbalanced datasets,a one-class sequence classification with adversarial network(OC-SCAN)model is proposed.In this method,autoencoder based on LSTM is used to transform the original online behavior sequence of students into a student representation vector.We use the sample of non-risk students to train the discriminator of boundary GAN model which is different from the traditional GAN model to predict risk students.In order to further improve the accuracy and stability of the model,we propose an integrated framework,en-OC-SCAN,to train multiple autoencoder with different structures independently.Experiments on real university curriculum datasets show that the performance of the proposed algorithm is better than other one classification and two classification algorithms on unbalanced datasets.(3)In order to improve the interpretability of deep learning model and carry out attribution analysis for risk students,an interpretable memory network(IMNet)is proposed.The model captures long-term behavior patterns through memory network module,and finds the most relevant sub-sequence segment of prediction decision through attention mechanism.In this way,the model makes the prediction easy to be understood by final users.Experiments on real university curriculum datasets show that this method can effectively predict risk students and find their potential factors.