Knowledge graphs have been widely used in recent years to build recommendation systems. However, in related research, the main focus has been on improving recommen dation algorithms, with less attention paid to the type of knowledge graph that is more conducive to k nowledge recommendations. This paper addresses knowledge recommendati on systems in the shipbuilding domain by constructing knowledge graphs in two ways and comparing their performance in knowledge recommendation. One type of knowledge graph is built based on classification tags of knowle dge documents, characterized by its simplicity and sparsity; the other is constructed automatically using machine learning, linking knowledge documents together based on concepts and relationships extracted by the algorithm, featuring complexity and density. To recommend shipbu ilding knowledge, a context-awa re mechanism was employed, gathering information from the user's task environment and linking it to the knowledge graph. Then, using RippleNet, the system spreads the user's interests within the k nowledge graph and infers the required knowledge documents. Experimental results show that the sparse knowledge graph achieved better recommendation results. We believe this is due to the human exp ert experience relied upon during the construction of the sparse knowle dge graph, namely a knowledge classification system oriented towards knowledge applications.