Research On Predictive Model Of College Students’ Academic Performance Based On Time Series Analysis

As the number of university enrollment in China increases year by year,the quality of university teaching has received widespread attentions from people from all walks of life.Academic performance is an important indicator to measure not only the effectiveness of students’ learning over a period of time,but also the level of education in colleges and universities.At the same time,a wide variety of colleges and universities have established digital campus environments with the rapid development of the Internet of Things,cloud computing and big data.In this process,massive amounts of student behavior data are generated and accumulated on digital campus.These data sets cover the entire process of students’ daily study and life,which have been drawn much attentions from both researchers and university administrators.However,in practical applications,restricted by factors such as technology,most studies on academic performance prediction are often carried out using a single data source,therefore,the prediction accuracy is often not ideal,and the deliverable feedback is relatively limited.To initially alleviate the above challenges,this article focuses on analyzing and mining the students’ behavioral change by means of multi-spatial data integration on digital campus,integrating multi-source data in Central China Normal University including(ⅰ)cloud classroom data,(ⅱ)smart card data,(ⅲ)WiFi data,(ⅳ)campus network usage data,and(ⅴ)digital educational administration data.Subsequently,multi-class classification model is built based on machine learning algorithms to predict students’ academic performance.On this basis,visualized feedback is provided to drive positive behavioral changes and improve students’academic performance.First of all,this study proposes a framework for academic performance prediction,including prediction model and early warning feedback model.With regards to the prediction model,the following four parts are included:① Behavior perception and extraction,on the one hand,the student information is encrypted to prevent the disclosure of student privacy;on the other hand,the educational administration system.all-in-one card,cloud classroom online learning and WiFi data are collected in the digital campus environment.For example,the all-in-one card data focuses on extracting information such as consumption time,consumption amount,and consumption location,while for cloud classroom data,it collects log-in and offline and emotional probability information.Based on matching the extracted data information with student behavior,construct a time series data set of students’ multi-dimensional behavior divided by week.②Feature extraction,Based on the time series data set,a quantitative calculation representation model of the behavior sequence is constructed,which calculates the change of behavior from four aspects.First,linear regression analysis is used to quantify the linear change characteristics of student behavior using linear indicators(inflection point,slope,residual,etc.);non-linear indicators(including:HMM entropy,Lee index,Hurst index,de-trend volatility analysis)Etc.)Quantify the non-linear change characteristics of student behavior;use deep learning algorithms and use long-term short-term memory model(LSTM)to extract the temporal change characteristics of student behavior.Further,in order to comprehensively consider the impact of different types of behavior indicators on academic performance,this research starts from the perspective of weighted fusion,calculates the "weighted characteristics" of various indicators,and systematically quantifies and analyzes the multi-dimensional behavior characteristics.The experimental results show that the behavioral features extracted in this paper can better describe the behavior patterns of students.③Intelligent prediction,Pearson correlation analysis is used to preliminarily screen the features that have a greater impact on academic performance,and on the basis of feature selection,machine learning algorithms such as XGBoost,SVM,KNN,GBRT,and RF are used to construct predictive models,and the classic models are compared.Forecast performance.The study found that the performance comparison shows that each prediction algorithm has good prediction performance on multiple data sets.As far as early warning feedback is concerned,this study uses a 95%confidence interval box diagram to visualize analysis and early warning from the perspective of linearity,nonlinearity,and timing of behavior patterns,It can be obtained from the box diagram.The characteristics of this research can be prepared to distinguish students in different grades.Secondly,the effectiveness of the proposed framework is verified on two data sets:① Academic performance prediction based on multi-source campus data(multi-source+small sample).It has been verified that the model established in our study can predict academic performance accurately,and at the same time early warning of academic performance for students who are at risk is provided.②Academic performance prediction based on online behavior data(single source+large sample).Taking into account that(ⅰ)in previous studies,there is often a lack of comparisons between different student groups;and(ⅱ)due to the constraints of objective conditions in practice,the sample size of multi-sources data is relatively limited,to some extent,the academic predictive model based on multi-spatial data fusion may not have universal applicability.Therefore,this study also establish an academic performance prediction model based on a single data source(campus network usage data),which can predict performance on a quite larger sample.The results show that there is a significant correlation between students’online behavior and their academic performance.In conclusion,the model established in this study can predict academic performance for college students with accuracy above 80%.The prediction model can also better distinguish whether students are at risk of failing courses,regardless of whether it is a multi-source small sample or a single-source large sample.In addition,for students with different liberal arts backgrounds,online behavior has different effects on their academic performance.