Residual convolutional graph neural network with subgraph attention pooling

The pooling operation is used in graph classification tasks to leverage hierarchical structures preserved data and reduce computational complexity. However, shrinkage discards details, existing methods may lead the loss of key features. In this work, we propose a residual convolutional neural network tackle problem features losing. Particularly, our contributions are threefold: (1) Different from methods, new strategy calculate sorting values verify their importance for classification. Our does not only use simple nodes but also neighbors accurate evaluation its importance. (2) We design layer architecture with connection. By feeding discarded back into architecture, probability losing critical (3) method graph-level representation. messages each node aggregated separately, then different attention levels assigned merged representation retain structural information experimental results show that leads state-of-the-art on multiple benchmarks.