Knowledge graph (KG) has been widely adopted in recommendation systems to alleviate sparsity and cold-start problems that generally exist in collaborative filtering-based methods. Such methods have improved performance to some extent. However, they only consider specific relations (interaction relations), which leads to low recommendation accuracy. This paper proposes a hierarchical KG attention network (HKGAN) to fuse multiple relations (interaction, social, and similarity) by decomposing rich semantic relations in the original KG. This model encodes different types of entities of a KG with a novel hierarchical attention mechanism to disseminate node-level and graph-level features. More accurate representations of users and items can be learned from multiple heterogeneous features. Then, the prediction scores are calculated by performing inner product operations on the representations of users and items. Experimental results on three public datasets show that the proposed model can significantly outperform state-of-the-art baselines.