University mathematics involves abstract concepts, rigorous logic, and complex interrelations among knowledge points. Traditional teaching methods are inadequate for capturing these connections, often resulting in fragmented knowledge, ambiguous reasoning paths, and poor knowledge transfer. To achieve accurate knowledge analysis and intelligent reasoning, this paper focuses on three core subjects—Calculus, Linear Algebra, and Probability Theory—and proposes an approach that integrates knowledge graphs, graph neural networks, and large language models for correlation and reasoning path mining. Specifically, a knowledge graph containing entities, relations, and attributes is constructed. The correlation strength between knowledge points is quantified from structural, semantic, and logical perspectives to identify both core and implicit associations. Subsequently, through dual-path graph reasoning combined with logical verification using a large language model, shortest, optimal, and multi-branch reasoning paths are generated. Experimental results demonstrate that the proposed method achieves higher accuracy and strong interpretability, providing theoretical support and technical reference for instructional design, personalized learning path planning, and intelligent education systems.