The goal of object navigation is to navigate an agent to a target object using visual input. Without GPS and the map, one challenge of this task is how to locate the target object in the unseen environment, especially when the target object is not in the field of view. Previous works use relation graphs to encode the concurrence relationships among all the object categories, but these relation graphs are usually too flat for the agent to locate the target object efficiently. In this paper, a Hier archical Graph Convolutional Neural Network (HGCNN) is proposed to encode the object relationships in a hierarchical manner. Specifically, the HGCNN consists of two graph convolution blocks and a graph pooling block, which constructs the hierarchical relation graph by learning an area-level graph from the object-level graph. Consequently, the HGCNN based framework enables the agent to locate the target object efficiently in the unseen environment. The proposed model is evaluated in the AI2-iTHOR environment, and the performance of object navigation shows a significant improvement.