Regarding the code for “CAG for IEEE Trains on Artificial Intelligence”

Complex networks exhibit inherent community structures that contain rich topology information of graphs. Existing graph neural networks (GNNs) have not fully utilized community structures and integrating them into GNNs has the potential to enhance nodes' representation capabilities and prediction performance. In this paper, we propose a community-aware graph neural network (CAG), which designs a community subgraph explorer (CSE) that leverages monte carlo tree search (MCTS) to select the most informative subgraphs within communities. These denoised subgraphs effectively capture and reflect the structural characteristics of the communities.

To discriminate node representations between communities, we introduce contrastive learning and a community-aware loss that facilitates the acquisition of structural similarities within communities while discerning differences among distinct communities. 

Extensive experiments demonstrate that CAG consistently improves GNNs' performance potential on node classification and achieves state-of-the-art accuracies (improved up to 3.9\%). which emphasizes the crucial role of community structures in designing and training GNNs.