A Novel Contrast Learning Model Based on Data Augmentation

Graph contrastive learning has shown remarkable advancements as a semi-supervised method for comprehensive graph representation learning. However, previous graph contrastive learning methods have shown insufficient exploration of underlying community attributes. Prior research often assumes that community attributes have a uniform impact on graph representations, which usually introduces unnecessary representational biases. In response to this obstacle, the paper introduces an ingenious graph feature augmentation and topology augmentation method. This approach scores nodes based on community attributes and creates an adjacency matrix based on node similarity. Furthermore, we propose an innovative contrastive learning approach that successfully retains the overall topological structure details by maximizing the mutual information between the graph and its feature embeddings. Extensive experimental validation shows that the proposed method outperforms existing methods in semi-supervised node classification tasks, with the highest accuracy increased by 0.7% and the highest F1 score increased by 1.32%, paving the way for further research in the realm of graph contrastive learning.