Exploring graph capsual network for graph classification

Graph Neural Network (GNN) has received tremendous attention due to their power in learning graph representations by modeling the topological structure and aggregating feature information. However, the scalar node representations learned from GNN may not be sufficient to effectively preserve the attributes of the node/graph features, resulting in sub-optimal graph representation. Repeated averaging gathers too much noise, which makes the features of nodes in different classes over-mixed and leads to the problem of over smoothing. Inspired by the concept of capsule network proposed by Hinton, we propose a new framework for graph classification, named CapsualGNN, which takes full advantage of Graph Neural Network and Capsule Network. Specifically, we firstly represent nodes as groups of capsules, in which each capsule extracts distinctive features of its corresponding node. Then, we exploit routing mechanism to capture important information and properties at the graph level by the generated multiple embeddings for each graph, and utilize attention mechanism to focus on important features. Finally, to solve the problem of over smoothing, we introduce class residual connection for GCN. In addition, we also introduce parameter for distinguishing self-connected nodes and other nodes. We evaluate the framework by using six graph datasets on biological information and social networks, and demonstrate that CapsualGNN outperforms other SOTA techniques on the task of graph classification.