The rapid growth in the number of scientific documents makes it increasingly necessary to provide researchers with efficient paper recommendation services. Since scientific documents contain many mathematical expressions and rich text information, effectively mining this information to improve recommendation quality is a challenging and necessary task. Therefore, this paper proposes a method that fuses content-based and Graph Convolutional Network (GCN) introducing the multi-head attention mechanism for scientific document recommendation. The method employs Hesitant Fuzzy Set (HFS) and Bidirectional Encoder Representations from Transformers (BERT) to calculate the similarity between mathematical expressions and their context, respectively, and utilizes GCN to learn the complex relationships between nodes and mine the potential similarity between nodes. Then, the multi-head attention mechanism is introduced to dynamically assign importance weights to neighbor nodes to enhance the expression of embedding vectors, thereby generating more accurate and personalized recommendation lists. Experimental results show that the method performs well on both Ntcir and extended Chinese datasets, achieving maximum F1_Score vales of 0.817 and 0.837 at Top-50 and Top-30, respectively.