Graph Neural Networks (GNNs) are an effective framework for graph representation learning in real-world appli- cations. However, despite their increasing success, they remain notoriously challenging to interpret, and their predictions are hard to explain. Nowadays, several recent works have proposed methods to explain the decisions made by GNNs. However, they only aggregate information from the same type of neighbors or indiscriminately treat homogeneous and heterogeneous neighbors similarly. Based on these observations, we propose HGExplainer, an explainer for heterogeneous GNNs to comprehensively capture structural, semantic, and attribute information from homogeneous and heterogeneous neighbors. We first train the GNN model to represent the predictions on a heterogeneous network. To make the explainable predictions, we design the model to capture heterogeneity information in calculating the joint mutual information maximization, extracting the meta-path-based graph sampling to generate more prosperous and more accurate explanations. Finally, we evaluate our explainable method on synthetic and real-life datasets and perform concrete case studies. Extensive results show that HGExplainer can provide inherent explanations while achieving high accuracy.