This study utilizes Knowledge Graph Attention Network (KGAT) to embed cultural heritage ontology data, thereby creating a recommender system based on the similarity of the heritages. We build a cultural heritage graph using the ontology data and embed all nodes and edges in the graph as vectors. We then make a recommendation result based on the cosine similarity. Also, we propose a method to embed a new cultural heritage using existing embedding vectors without retraining the model, and hence effectively addressing the cold start problem. Experiments demonstrate that our system creates a list of relevant cultural heritages as recommendation and handle new items efficiently without additional training. This method offers a new perspective on cultural heritage, with potential future research integrating user information for more precise recommendations.