and practicalsignificance. How to model learner knowledge from learner answer records has always been the research focus in exerciserecommendation under sparse data. Therefore,an exercise recommendation algorithm combining deep knowledge tracking and knowledgematrix complementation is proposed to address the problems of sparse data and ignoring group features in existing exerciserecommendation methods, which is divided into two modules: knowledge?
level modeling and knowledge matrix complementation.Firstly,the learner knowledge level matrix is obtained through?
the training of deep knowledge tracking model. By this way,it realizes theknowledge level modeling of learners and accurately explores the mastery level of learners’ knowledge concepts. Secondly,the knowledgelevel of similar users is fused by considering the near - neighbor information of learners and using the group features among learners.Finally,
the matrix decomposition module is introduced to perform the knowledge matrix complementation. It can predict the score oflearners’ undone exercises and has the advantage of alleviating the data sparsity problem. The recommendation algorithm takes into consideration both the group commonality of learners and the sparse matrix of learners’ knowledge levels. The proposed algorithm effectivelyimproves the accuracy, recall and F1 value of the recommendation results compared with the other algorithm, and its performanceadvantage becomes more obvious as the number of exercises recommended increases.