ComMGAE: Community Aware Masked Graph AutoEncoder

Independent of graph augmentation techniques, graph autoencoders (GAEs) have yielded promising results in the realm of self-supervised learning. However, GAEs tend to over-emphasize proximity information at the expense of structural information, leading to relatively poor performance on some downstream tasks such as node classification. To address this issue, we propose a novel GAE framework via community, named Community aware Masked Graph AutoEncoder(ComMGAE). Since community represents a high-order structure of a graph, characterized by a group of densely connected nodes, ComMGAE can import structural information from community semantics to the self-supervised generative learning on the graph. Firstly, we partition the community structure by using the community detection algorithm and calculate the community strength. Then we introduce a novel community-guided graph masking strategy to learn more about the graph structure during the encoding process. Moreover, we mask and reconstruct both the structure and attribution of the graph and employ a graph neural network as the decoder to enrich learning representations with compressed information. Finally, experimental results on node classification demonstrate that the proposed ComMGAE preserves both the graph topology and semantic information effectively and outperforms other state-of-the-art baselines on a series of benchmarks.