A Dual Adaptive PageRank Graph Neural Network with Structural Augmentation

With the gradual development of deep learning-related technologies, Graph Neural Networks (GNNs) have achieved great success in graph representation learning, promoted the development of various tasks on graphs, and solved many problems in the real world, for example, demonstrated strong performance in social networks, recommendation systems, and connection prediction, the core of which is to update node representations by aggregating nearest-neighbor information. However, most of the existing GNN designs are used in homophilic graphs, where nodes of the same class tend to be more easily connected, and do not perform well in heterophilic graphs, where the connected nodes tend to different classes, and the aggregation of nodes of different classes results in nodes without clear classes. Existing methods suggest capturing some useful information in heterophilic graphs by expanding the receptive field of the GNN. However, expanding the receptive fields of nodes in heterophilic graphs can also lead to "over-smoothing" between nodes, which makes nodes indistinguishable from each other and leads to a decrease in classification accuracy. To address the above problem, we propose a novel decoupled GNN model called Dual Adaptive PageRank Graph Neural Network with Structural Augmentation (DAPRGNN) in this work. To be specific, we add a node adaptation operation based on Generalized PageRank to the propagation, called Dual Adaptive PageRank (DAPR). DAPR enables the model to personalize the selection of neighboring node information, which improves the capture of informationrich nodes, especially in heterophilic graphs. In addition, in order to achieve better performance in Deep Graph Neural Networks, we also randomly remove certain edges in the input graph at each training stage to reduce the effect of "over-smoothing" on the final classification results. Through extensive experiments on representative real-world homophilic and heterophilic datasets, DAPRGNN outperforms the popular baseline GNN model in semi-supervised node classification tasks, and achieves better classification results.