Personalized Exercise Recommendation Combined With Deep Knowledge Tracing

Personalized exercise recommendation is an important topic in the field of educational data mining.The exercise recommendation algorithm can model the student's learning progress according to the student's exercise submission records,and recommend the appropriate exercises to the student according to certain rules to achieve the purpose of assisting students in online learning.It is of great significance that research on personalized exercise recommendation can effectively improve the quality of online education.At present,further development of exercise recommendation algorithms faces two main problems:(1)Existing exercise recommendation methods are mainly divided into algorithms based on collaborative filtering and algorithms based on knowledge modeling.The method of modeling knowledge starts from the personal level of students,and most of them ignore the use of common features between similar students;while the collaborative filtering method mines the data of students' questions,finds similar users and exercises,and often ignores student's knowledge status.(2)Existing exercise recommendation strategies are mainly for scoring and ranking resources or relying on experts to formulate them.The efficiency is low,and it is difficult to adjust the recommendation direction in time according to changes in students' knowledge levels.In response to the above exercises,this paper proposes a personalized exercise recommendation method combining deep knowledge tracing and collaborative filtering method.The method first uses deep knowledge tracing to model student knowledge,and then combined with collaborative filtering method to calculate the correct probability of students' exercises,and based on this probability,recommend the exercises within a certain difficulty range to students.This method also refers to the personal knowledge level and the neighbor information of students in similar situations,has better model accuracy,and can recommend suitable content according to the difficulty range.And verified by experiments based on real data,the specific contributions of this article are as follows:1.In order to improve the accuracy of the exercise recommendation algorithm model,innovatively combine the deep knowledge tracing model with the collaborative filtering algorithm,and design a new exercise recommendation model DKT-CF to improve the accuracy of exercise recommendation algorithm.First,based on the historical data submitted by the students,the personal knowledge level matrix of the students is output through the deep knowledge tracing model DKT,and the matrix is used to replace the exercise-score matrix.The collaborative filtering algorithm is used to find similar student groups,and the neighbor information is introduced to generate personalized exercise recommendations.Experiments proved that the model significantly improves the accuracy of the recommendation algorithm by combining the students'individual knowledge level information and the group information of similar students.Compared with DKT+,the improved model of DKT and DKT,the precision increases by 2.0%and 2.2%respectively;the recall increases by 3.1%and 0.6%respectively;the F1 score increases by 0.04 and 0.01 respectively.2.By introducing the difficulty range parameter,we can recommend questions with the correct difficulty to students based on the prediction to student's knowledge status,so that the recommendation results are more flexible and interpretable than simple ranking recommendation strategies.The model is used as a student simulator to improve the exercise recommendation strategy.By simulating the follow-up submission sequence of different difficulty recommendation questions for students to calculate the changes of students'knowledge level,the algorithm can adaptively find the appropriate difficulty that can increase the student's knowledge level at most.Experiments on two real data sets prove that by selecting the difficulty that makes the most student's knowledge improve,the rate of knowledge level increase is equivalent to 1.57 times and 2.12 times than the recommendation strategy for resource score ranking,and no additional expert knowledge is required.