The paper introduces HetGNN, a heterogeneous graph neural network model designed to address the challenges of representing nodes in heterogeneous graphs. These challenges include incorporating heterogeneous structural information and content information, such as text or images, associated with each node. The model uses a random walk with restart strategy to sample strongly correlated heterogeneous neighbors for each node and groups them based on node types. It then employs two neural network modules to aggregate the feature information of these neighbors. The first module encodes deep feature interactions of heterogeneous contents, while the second module aggregates content embeddings of different neighboring groups and combines them using an attention mechanism to consider the impacts of different node types. The model is trained using a graph context loss and mini-batch gradient descent. Extensive experiments on various datasets demonstrate that HetGNN outperforms state-of-the-art baselines in tasks such as link prediction, recommendation, node classification, and clustering. The paper also includes an analysis of the model's performance and hyper-parameter sensitivity.The paper introduces HetGNN, a heterogeneous graph neural network model designed to address the challenges of representing nodes in heterogeneous graphs. These challenges include incorporating heterogeneous structural information and content information, such as text or images, associated with each node. The model uses a random walk with restart strategy to sample strongly correlated heterogeneous neighbors for each node and groups them based on node types. It then employs two neural network modules to aggregate the feature information of these neighbors. The first module encodes deep feature interactions of heterogeneous contents, while the second module aggregates content embeddings of different neighboring groups and combines them using an attention mechanism to consider the impacts of different node types. The model is trained using a graph context loss and mini-batch gradient descent. Extensive experiments on various datasets demonstrate that HetGNN outperforms state-of-the-art baselines in tasks such as link prediction, recommendation, node classification, and clustering. The paper also includes an analysis of the model's performance and hyper-parameter sensitivity.