Deep Graph Infomax (DGI) is a novel approach for unsupervised learning of node representations in graph-structured data. DGI maximizes mutual information between patch representations and high-level summaries of graphs, derived using graph convolutional network architectures. This method allows for the reuse of learned patch representations for downstream node-wise learning tasks. Unlike most prior approaches that rely on random walk objectives, DGI is applicable to both transductive and inductive learning setups. The paper demonstrates competitive performance on various node classification benchmarks, sometimes even surpassing the performance of supervised learning methods. The key contributions include the introduction of DGI, its theoretical motivation, and extensive experimental validation on multiple datasets.Deep Graph Infomax (DGI) is a novel approach for unsupervised learning of node representations in graph-structured data. DGI maximizes mutual information between patch representations and high-level summaries of graphs, derived using graph convolutional network architectures. This method allows for the reuse of learned patch representations for downstream node-wise learning tasks. Unlike most prior approaches that rely on random walk objectives, DGI is applicable to both transductive and inductive learning setups. The paper demonstrates competitive performance on various node classification benchmarks, sometimes even surpassing the performance of supervised learning methods. The key contributions include the introduction of DGI, its theoretical motivation, and extensive experimental validation on multiple datasets.