GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training This paper proposes GCC, a self-supervised graph neural network pre-training framework that captures universal network topological properties across multiple graphs. The pre-training task is designed as subgraph instance discrimination, leveraging contrastive learning to enable graph neural networks to learn intrinsic and transferable structural representations. Extensive experiments on three graph learning tasks and ten graph datasets show that GCC pre-trained on diverse datasets achieves competitive or better performance compared to task-specific and scratch-trained models. This suggests that pre-training and fine-tuning paradigms have great potential for graph representation learning. GCC is designed to learn structural representations across graphs by using contrastive learning to design the graph pre-training task as instance discrimination. The framework uses subgraph instance discrimination as the pre-training task, aiming to distinguish vertices based on their local structures. For each vertex, subgraphs are sampled from its multi-hop ego network as instances. A graph neural network (specifically, the GIN model) is used to map the underlying structural patterns to latent representations. The graph encoder is forced to capture universal patterns across different input graphs. GCC is compared with existing methods such as DeepWalk, Struc2vec, and GraphWave. The results show that GCC achieves competitive or better results than these models on tasks like node classification. The framework is also compared with InfoGraph and other recent approaches, showing that GCC can offer comparable or superior performance on out-of-domain tasks. GCC is evaluated on three graph learning tasks: node classification, graph classification, and similarity search. The results show that GCC achieves competitive performance on these tasks, demonstrating the effectiveness of the pre-training and fine-tuning paradigm. The framework is also compared with other methods such as DGCNN and GIN, showing that GCC can achieve better performance on some tasks. The paper also presents ablation studies on the effect of pre-training, contrastive loss mechanisms, and pre-training datasets. The results show that pre-training provides a better starting point for fine-tuning than random initialization. The contrastive loss mechanisms show that MoCo has stronger expression power than E2E, but the effect of a large dictionary size is not as significant as in computer vision tasks. The pre-training datasets show that using more datasets leads to higher accuracy on both US-Airport and COLLAB datasets. Overall, the paper demonstrates that a graph neural network encoder pre-trained on several popular graph datasets can be directly adapted to new graph datasets and unseen graph learning tasks. The results show that the pre-trained model achieves competitive and sometimes better performance compared to models trained from scratch, demonstrating the transferability of graph structural patterns and the effectiveness of the GCC framework in capturing these patterns.GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training This paper proposes GCC, a self-supervised graph neural network pre-training framework that captures universal network topological properties across multiple graphs. The pre-training task is designed as subgraph instance discrimination, leveraging contrastive learning to enable graph neural networks to learn intrinsic and transferable structural representations. Extensive experiments on three graph learning tasks and ten graph datasets show that GCC pre-trained on diverse datasets achieves competitive or better performance compared to task-specific and scratch-trained models. This suggests that pre-training and fine-tuning paradigms have great potential for graph representation learning. GCC is designed to learn structural representations across graphs by using contrastive learning to design the graph pre-training task as instance discrimination. The framework uses subgraph instance discrimination as the pre-training task, aiming to distinguish vertices based on their local structures. For each vertex, subgraphs are sampled from its multi-hop ego network as instances. A graph neural network (specifically, the GIN model) is used to map the underlying structural patterns to latent representations. The graph encoder is forced to capture universal patterns across different input graphs. GCC is compared with existing methods such as DeepWalk, Struc2vec, and GraphWave. The results show that GCC achieves competitive or better results than these models on tasks like node classification. The framework is also compared with InfoGraph and other recent approaches, showing that GCC can offer comparable or superior performance on out-of-domain tasks. GCC is evaluated on three graph learning tasks: node classification, graph classification, and similarity search. The results show that GCC achieves competitive performance on these tasks, demonstrating the effectiveness of the pre-training and fine-tuning paradigm. The framework is also compared with other methods such as DGCNN and GIN, showing that GCC can achieve better performance on some tasks. The paper also presents ablation studies on the effect of pre-training, contrastive loss mechanisms, and pre-training datasets. The results show that pre-training provides a better starting point for fine-tuning than random initialization. The contrastive loss mechanisms show that MoCo has stronger expression power than E2E, but the effect of a large dictionary size is not as significant as in computer vision tasks. The pre-training datasets show that using more datasets leads to higher accuracy on both US-Airport and COLLAB datasets. Overall, the paper demonstrates that a graph neural network encoder pre-trained on several popular graph datasets can be directly adapted to new graph datasets and unseen graph learning tasks. The results show that the pre-trained model achieves competitive and sometimes better performance compared to models trained from scratch, demonstrating the transferability of graph structural patterns and the effectiveness of the GCC framework in capturing these patterns.