Graph Transformer Networks (GTNs) are proposed to learn node representations on heterogeneous graphs by generating new graph structures. Unlike traditional GNNs that assume fixed and homogeneous graphs, GTNs can identify useful connections between unconnected nodes and learn effective node representations on new graphs in an end-to-end manner. The core layer of GTNs, the Graph Transformer layer, learns soft selection of edge types and composite relations to generate multi-hop connections, known as meta-paths. GTNs generate new graph structures based on data and tasks without domain knowledge, achieving the best performance in three benchmark node classification tasks without requiring predefined meta-paths. GTNs can adaptively generate variable-length meta-paths and provide interpretable insights into effective meta-paths for prediction. The framework is capable of learning new graph structures and node representations through convolution on the generated graphs, leading to state-of-the-art performance on heterogeneous graphs. GTNs are shown to be effective in learning node representations and adaptively generate meta-paths based on the dataset, outperforming existing methods in node classification tasks. The model's ability to learn and adaptively generate meta-paths makes it a powerful approach for node classification on heterogeneous graphs.Graph Transformer Networks (GTNs) are proposed to learn node representations on heterogeneous graphs by generating new graph structures. Unlike traditional GNNs that assume fixed and homogeneous graphs, GTNs can identify useful connections between unconnected nodes and learn effective node representations on new graphs in an end-to-end manner. The core layer of GTNs, the Graph Transformer layer, learns soft selection of edge types and composite relations to generate multi-hop connections, known as meta-paths. GTNs generate new graph structures based on data and tasks without domain knowledge, achieving the best performance in three benchmark node classification tasks without requiring predefined meta-paths. GTNs can adaptively generate variable-length meta-paths and provide interpretable insights into effective meta-paths for prediction. The framework is capable of learning new graph structures and node representations through convolution on the generated graphs, leading to state-of-the-art performance on heterogeneous graphs. GTNs are shown to be effective in learning node representations and adaptively generate meta-paths based on the dataset, outperforming existing methods in node classification tasks. The model's ability to learn and adaptively generate meta-paths makes it a powerful approach for node classification on heterogeneous graphs.