This paper proposes HetGNN, a heterogeneous graph neural network model, to address the challenges of representation learning in heterogeneous graphs (HetG). HetG contains multiple types of nodes and edges, along with heterogeneous attributes or contents (e.g., text or image) associated with each node. Existing methods struggle to effectively incorporate both structural and content heterogeneity. HetGNN introduces a random walk with restart strategy to sample fixed-size strongly correlated heterogeneous neighbors for each node and groups them by node type. It then designs a neural network architecture with two modules: the first encodes "deep" feature interactions of heterogeneous contents to generate content embeddings, while the second aggregates content embeddings of different neighboring groups and combines them using an attention mechanism to obtain the ultimate node embedding. The model is trained using a graph context loss and mini-batch gradient descent. Extensive experiments on several datasets demonstrate that HetGNN outperforms state-of-the-art baselines in various graph mining tasks, including link prediction, recommendation, node classification, and clustering. The main contributions include formalizing the problem of heterogeneous graph representation learning, proposing HetGNN, and conducting extensive experiments to validate its effectiveness.This paper proposes HetGNN, a heterogeneous graph neural network model, to address the challenges of representation learning in heterogeneous graphs (HetG). HetG contains multiple types of nodes and edges, along with heterogeneous attributes or contents (e.g., text or image) associated with each node. Existing methods struggle to effectively incorporate both structural and content heterogeneity. HetGNN introduces a random walk with restart strategy to sample fixed-size strongly correlated heterogeneous neighbors for each node and groups them by node type. It then designs a neural network architecture with two modules: the first encodes "deep" feature interactions of heterogeneous contents to generate content embeddings, while the second aggregates content embeddings of different neighboring groups and combines them using an attention mechanism to obtain the ultimate node embedding. The model is trained using a graph context loss and mini-batch gradient descent. Extensive experiments on several datasets demonstrate that HetGNN outperforms state-of-the-art baselines in various graph mining tasks, including link prediction, recommendation, node classification, and clustering. The main contributions include formalizing the problem of heterogeneous graph representation learning, proposing HetGNN, and conducting extensive experiments to validate its effectiveness.