Structured entities analysis is the basis of the modern science, such as chemical science, biological science, environmental science and medical science. Recently, a huge amount of computational models have been proposed to analyze structured entities such as chemical molecules and proteins. However, the problem becomes complex when local structural entity graphs and a global entity interaction graph are both involved. The unique graph of graphs structure cannot be properly exploited by most existing works for structural entity analysis. Some works that build neural networks on the graph of graphs cannot preserve the local graph structure effectively, hence, reducing the expressive power of the model. In this paper, we propose a Powerful Graph Of graphs neural Network, namely PGON, which has 3-Weisfeiler-Lehman expressive power and captures the attributes and structural information from both structured entity graphs and entity interaction graph hierarchically. Extensive experiments are conducted on real-world datasets, which show that PGON outperforms other state-of-the-art methods on both graph classification and graph interaction prediction tasks.

结构实体分析是现代科学的基础，如化学科学、生物科学、环境科学和医学科学。最近，已经提出了大量的计算模型来分析结构化实体，例如化学分子和蛋白质。然而，当同时涉及局部结构实体图和全局实体交互图时，问题变得复杂。大多数现有的结构实体分析工作无法正确利用图结构的独特图。一些在图的图上构建神经网络的工作不能有效地保留局部图结构，因此降低了模型的表达能力。在本文中，我们提出了一个强大的图神经网络，即 PGON，它具有 3-Weisfeiler-Lehman 表达能力，并从结构化实体图和实体交互图中分层捕获属性和结构信息。在真实世界的数据集上进行了大量实验，结果表明 PGON 在图分类和图交互预测任务上都优于其他最先进的方法。