Graph Transformer Networks (GTNs) are proposed to address the limitations of existing Graph Neural Networks (GNNs) in handling heterogeneous graphs. GNNs typically assume fixed and homogeneous graph structures, which can lead to ineffective convolution and poor performance on noisy or heterogeneous graphs. GTNs learn to transform heterogeneous input graphs into new graph structures defined by meta-paths, while simultaneously learning effective node representations. The core of GTNs is the Graph Transformer layer, which soft selects edge types and composite relations to generate multi-hop connections, or *meta-paths*. Experiments show that GTNs can learn new graph structures based on data and tasks without domain knowledge, achieving state-of-the-art performance in node classification tasks on heterogeneous graphs. The contributions of GTNs include a novel framework for learning new graph structures, interpretable graph generation, and superior performance compared to state-of-the-art methods that require pre-defined meta-paths.Graph Transformer Networks (GTNs) are proposed to address the limitations of existing Graph Neural Networks (GNNs) in handling heterogeneous graphs. GNNs typically assume fixed and homogeneous graph structures, which can lead to ineffective convolution and poor performance on noisy or heterogeneous graphs. GTNs learn to transform heterogeneous input graphs into new graph structures defined by meta-paths, while simultaneously learning effective node representations. The core of GTNs is the Graph Transformer layer, which soft selects edge types and composite relations to generate multi-hop connections, or *meta-paths*. Experiments show that GTNs can learn new graph structures based on data and tasks without domain knowledge, achieving state-of-the-art performance in node classification tasks on heterogeneous graphs. The contributions of GTNs include a novel framework for learning new graph structures, interpretable graph generation, and superior performance compared to state-of-the-art methods that require pre-defined meta-paths.