<p>Due to the differences in learners’ knowledge mastery and learning needs, knowledge matching and appropriate exercise difficulty have always been important challenges faced by personalized exercise recommendation research. Furthermore, the logical relation among knowledge concepts is also an important factor that affects learners’ learning efficiency in exercise recommendation. However, most existing exercise recommendation researches only consider learners’ knowledge mastery, without taking into account their personalized feelings toward exercise difficulty. So, it is difficult to achieve precise recommendation for learners. To address these issues, this paper proposes a personalized exercise recommendation based on <b>K</b>nowledge <b>S</b>tructure and learners’ <b>A</b>ttempting <b>P</b>references (KSAP). Firstly, the method uses deep knowledge tracking (DKT) to obtain learners’ current knowledge state based on their interaction records, which is used to obtain learners’ weak knowledge concepts set. Secondly, Apriori association rules are used to mine the knowledge structure and generate knowledge map, which is further improved based on learners’ knowledge mastery to expand and rank their weak knowledge concepts set in order to obtain the corresponding sequence. Finally, in order to recommend the right difficulty level of exercises for learners, deep matrix factorization (DMF) is used to explore the high-order relation between learners and exercises to obtain their implicit exercise preferences. Their personalized explicit exercise preferences are mined from their historical attempting exercise data through statistical analysis. Then personalized exercises are recommended for learners by above three steps. In the experiments, the superiority of KSAP is verified by conducting comparative experiments with baseline methods. Moreover, the interpretability of KSAP is improved by considering the knowledge structure.</p>