Practical teaching is essential for talent development in higher education. However, the absence of an effective mentoring mechanism between senior and junior students, coupled with the underutilization of existing learning resources, results in low teaching efficiency and limited teacher engagement. To address these issues, a knowledge graph construction method based on the “Four-Step” Model for Practice-Oriented Teaching is proposed. By extracting key elements—such as practical projects, courses, and knowledge points—from diverse educational sources, a heterogeneous graph is constructed that represents multiple entity types and semantic relationships. HGNN is then used for node representation learning. The HGNN integrates relation-aware message passing and attention mechanisms, improving its ability to model complex teaching relationships. Experimental results show that the proposed method outperforms traditional graph neural networks in accuracy, F1 score, and AUC, demonstrating strong representation ability and practical value. HGNN provides a new technical approach for digitally modeling practical teaching in higher education.