A Component-level Attention based Adaptive Graph Convolutional Network

• DOI: 10.23919/ccc55666.2022.9901954
• Publication Date: 2022-10-11T19:33:35Z
• Authors: Xin Li; Yuhan Zhang; Wei Lu; Pan Zhu
• Keywords: Component (thermodynamics); Connected component; Smoothing

Graph neural networks have attracted more and attention in the task of learning node representations. Currently, most graph are usually applied to assortative graphs. They can't perform well disassortative graphs, where adjacent nodes tend different class labels. However, previous studies shown that it is effective for gather information from their neighbors not only graphs but also Based on this insight, paper proposes a novel Component-level Attention Adaptive Convolutional Network (CAAGCN) collect efficiently. Firstly, model collects dissimilar features among by introducing component-level mechanism. One component learns importance neighbor other regulates proportion similar characteristics nodes. Secondly, order between effectively, we optimize input function during process preprocess features. It can effectively alleviate over-smoothing. Finally, extensive experiments six verify effectiveness method classification task.