<p>In large-scale e-learning learning platforms, precisely assessing learners’ knowledge states is crucial for achieving personalized exercise recommendation. However, most existing studies only rely on answer accuracy to evaluate knowledge states, ignoring the cognitive abilities implied in answering behaviors. Meanwhile, the impact of learning ability on exercise recommendation has received little attention. Additionally, traditional knowledge graphs rely on preset relationships for topological evolution, making it difficult to capture deep semantics, and their dynamic evolution also poses challenges to graph embedding. Based on the above motivations, we propose an exercise recommendation approach via multi-behavior interaction perception and semantic-enhanced knowledge graph (MBSE). First, a time-sensitive multi-behavior conjugate knowledge tracing model is constructed: a GRU with behavior focus-temporal decay dual attention is used to analyze the evolution of learning behaviors, and a conjugate deep cross-network is employed to build behavior synergy/opposition coupling interactions to achieve accurate evaluation of knowledge states. Second, a dual-dimensional learner ability assessment framework is proposed to evaluate problem-solving ability and knowledge-acquisition ability. Third, an incremental progressive relational aggregation graph convolutional network is proposed to progressively represent relationships in knowledge graph, achieving deep semantic modeling of knowledge graph. Finally, a multi-task balanced recommendation model is constructed to realize accurate exercise selection and recommendation. Experimental results show that, compared with the second-best model, MBSE achieved maximum improvements of 1.35%, 1.15%, and 2.39% in Precision, Recall, and F1 respectively, with comparable Diversity, on the ASSIST2012-13 and ASSIST2009-10 datasets. These improvements indicate higher recommendation quality, helping reduce redundant practice, accelerate mastery, and improve learning outcomes.</p>