Exercise Prediction Algorithm

Exercise prediction refers to the use of student historical answer data to predict the correctness of students' unanswered exercises.It can show students their knowledge mastery status and facilitate students to check and fill vacancies;exercise prediction tasks can also help teachers in individualized teaching and teach students in accordance with their aptitude.This thesis divides the students' exercises into "practice exercises" and "exam exercises"."Practice exercises" refer to exercises for students to practice independently in order to improve their knowledge in the usual learning process."Exam exercises" refers to the exercises that all students are required to complete the exam within the specified time during the semester test.In the process of students' autonomous learning,due to the large number of "practice exercises",this type of exercises often shows that the number of intersections of the "practice exercises" answered by different students is small and the same "practice exercises" may be answered multiple times,that is,the data is sparse.On the contrary,when students answer "exam exercises",the "exam exercises" answered by different students at the same time are exactly the same,and each "exam exercise" can only be answered once.Therefore,in view of the differences between the two types of exercises,this thesis proposes exercise prediction methods for practice exercises and exam exercises respectively,which show good prediction effects on the public datasets.The research content of this thesis is as follows:(1)Aiming at practice exercises,this thesis proposes "knowledge tracing algorithm based on student behavior and exercise attribute".This algorithm not only makes full use of the characteristics of student behavior,but also considers the characteristics of exercise attribute,embeds the information of student behavior and exercise attribute into the dynamic key-value memory network,and models the students' exercise process to realize exercise prediction.Finally,good results were obtained on the ASSISTments2009-2010?updated dataset.(2)Aiming at exam exercises,this thesis proposes " A Weighting-based Student Exercise Matrix Factorization Model".The algorithm first obtains the student ability and the exercise difficulty through cognitive diagnosis models and data mining methods,which are incorporated into the loss function as prior knowledge.Secondly,the element-level alternating least squares optimization strategy is applied to speed up the update.Finally,experiments are designed on three datasets of Frc Sub,Math1 and Math2,which are better than the existing models in indicators such as MAE/RMSE,ACC and time complexity.(3)Aiming at the "new student-cold start" problem,this thesis proposes an incremental exercise prediction algorithm called "incremental update mechanism based on sequential meta-learning".The mechanism combines the records of "old students" and "new students" to refine "current knowledge" and predict the future answering conditions of "new students".At the same time,an incremental update mechanism is used to save retraining time.Finally,experiments were conducted on three public datasets of Frc Sub,Math1 and Math2,and the experimental results proved the effectiveness of the mechanism.In summary,the algorithms proposed in this thesis can comprehensively and accurately model the knowledge state of students answering practice exercises and exam exercises,thereby predicting the future answering situation,which is of great significance for students to improve their own knowledge state and promote the development of intelligent education.