Feature-enhanced community-guided graph contrastive learning

Contrastive learning, popular in computer vision, extracts shared information from different views. Recently, it has been introduced into graph networks, known as graph contrastive learning, to learn representations from graphs with scarce labels. However, directly applying contrastive schemes from computer vision to graph data is unreasonable. Randomly deleting edges, nodes, or attributes to generate augmented graph views results in low-quality, time-consuming, and unstable samples. Node features contain significant structural information, yet few studies focus on mining these structural views. Traditional node-level contrastive schemes may incorrectly classify negative samples as positive, or vice versa. To address these issues, this paper abandons data augmentation, using KNN to construct feature views and mine graph structure information. A community-guided graph contrast loss dynamically divides node communities during learning, ensuring reasonable sample division. A structure-aware module is also designed to focus on global graph structure information during contrastive learning. Experiments show this model achieves state-of-the-art results on multiple downstream tasks.