Exercise Recommendation Algorithm Based On Implicit Skill And Weighted Matrix Factorization

Exercises are important learning resources in the field of education.The personalized exercises recommendation algorithms play an important role in the online education platform.The cognitive diagnosis model is the most popular method of recommendation in the field of education,which describes the relationship between students,skills and exercises well.And it has been used for personalized exercises recommendation,which obtains the prediction results by modeling the knowledge level of students' skill.However,this method still has some defects when recommends exercises for students.Firstly,the accuracy of exercises recommendation of the cognitive diagnosis model mainly depends on the exactitude of the annotation of Q matrix,which is a mapping relationship matrix between exercises and skills,generally manually labeled by domain experts,and often inaccurate.Secondly,there are great quantity of exercises on the online education platform without Q matrix annotation,and the development of recommendation algorithms for which still has a long way to go.To solve the above two problems,two different personalized exercises recommendation algorithms are proposed respectively by this thesis and the experimental results on the public dataset are better than the existing algorithms,based on which,a prototype system for online teaching is designed and constructed.The specific research contents and innovations are as follows:(1)An exercise recommendation algorithm based on implied skills is proposed,which automatically obtains the relationship matrix between exercises and implied skills through knowledge tracing model at first.And then rebuilding the Q matrix by combining this relationship matrix and the Q matrix annotated by domain experts.Lastly,using this new Q matrix as input to the cognitive diagnosis models for exercise recommendation.(2)An exercise recommendation algorithm based on student-exercise weighted matrix factorization is proposed,which only need the exercise difficulty and student ability value as matrix factorization model's prior knowledge of unknown data instead of Q matrix.Moreover,the fast element-wise Alternating Least Squares update strategy has been applied to the proposed model.(3)A prototype system for online teaching are designed and constructed based on the above two exercise recommendation algorithms,which utilizes the student's test record for testing whether the student has mastered the knowledge,and recommends appropriate exercises for students to avoid repeating the exercises that have been mastered.The experimental results express that the F1 of two proposed personalized exercise recommendation algorithms on the public dataset achieved good results.These two algorithms can alleviate the error caused by manually labeled Q matrix,as well as provide feasible recommendation strategy for exercises without Q matrix annotation.Furthermore,the successful application of these two algorithms on the online teaching platform is conducive to improving the learning efficiency of students and the teaching effect of teachers.