Research On Student Performance Prediction Based On Campus Behavior Perception

With the advent of the big data era,colleges have accumulated a wealth of education and teaching information data under the continuous promotion of campus information construction.How to provide support and guidance for college education and teaching through massive educational data resources has become an urgent problem to be solved.Educational data mining(EDM)came into being in this context.EDM is an interdisciplinary research field that comprehensively uses the theories and technologies of pedagogy,psychology,statistics and computer science to solve the problems in educational research and teaching practice.As one of the most important research topics in this field,student performance prediction aims to use various kinds of relevant information of students to predict their future academic performance.Through the technology of student performance prediction,teachers can make use of limited teaching resources to teach students according to their aptitude and develop differentiated teaching methods for students.At the same time,colleges can also carry out early academic warning work for some students who have the risk of failing courses,repeating grades or even dropping out through the student performance prediction technology,and finally achieve the goal of talent training.Therefore,it is of great significance to study an efficient and accurate method of student performance prediction for improving teaching output and strengthening student management.Due to the popularization of campus credit card system in colleges,massive amounts of student behavior data on campus learning and life are recorded in a completely hidden way through campus smart cards.Considering that human behavior is closely related to learning ability,making full use of these data will help solve the problem of student performance prediction.Therefore,we analyze students’ daily campus behavior trajectory(i.e.the campus smart card records)to solve the problem of student performance prediction.Specifically,each student’ s campus smart card is recorded in terms of dimensions of date,time slot and campus location,so that each student’s behavior trajectory is encoded as a third-order tensor.Different from previous studies,this paper proposes a novel end-to-end tri-branch convolutional neural network(TB-CNN)architecture,which equipped with row-wise,column-wise and depth-wise convolution operations.After each convolution operation,an attention module is added to capture the characteristics of persistence,regularity and spatial perception of student behavior.In addition,we regard the prediction of students of different majors as a separate task,and further improve the accuracy of the prediction method by introducing the multi task learning strategy based on hard parameter sharing and the cost-sensitive loss function based on students’ poor ranking.We have carried out extensive experiments on large-scale real-world datasets,and verified the effectiveness of the proposed method by visualizing the attention heat map.In addition,we study the performance prediction method that focus on academically at-risk students.The primary goal of student performance prediction is to detect students who are at academic risk as early as possible and provide guidance and assistance in time.However,the goal of existing work is often to improve the overall predictive performance of students.To solve this problem,we sort the students in order of academic performance from low to high,and define the top-k students as the academically at-risk students with the worst academic performance.Therefore,we cast the academic performance prediction as a top-k ranking problem.Following the idea of cost sensitive learning,we propose a top-k focused loss function,which allocates a higher weight to the training sample in the higher position,and a smaller weight to the training sample behind the k-th position,so as to ensure the correct recognition of the academically at-risk students.Through the visualization of students’ feature embedding and detailed comparative experiments,we prove the excellent learning ability of the proposed method for students’ feature representation and the superiority of the proposed method compared with other recently proposed student performance prediction methods.