Curriculum-guided graph self-augmentation: A progressive deepening framework for GNNs.

Graph Neural Networks (GNNs) have become a powerful tool for modeling complex graph-structured data, achieving remarkable success in various graph mining tasks. However, many GNNs are limited to shallow architectures due to over-smoothing, which prevents them from effectively capturing long-range relationships between nodes. One of the contributing factors to over-smoothing is the use of uniform aggregation weights, indicating that the graph data lack sufficient information to support deep GNNs effectively. To address this challenge, we propose Curriculum-Guided Graph Self-Augmentation (CGGSA), a progressive deepening framework for GNNs. The training process begins with learning simple node representations aggregated from low-order neighborhoods, extracting effective guidance to augment both the graph structure and node features, thereby facilitating deeper information aggregation. The aggregation depth is then gradually increased to capture more complex high-order neighborhood dependencies, enabling the model to learn long-range interactions and improve its performance. Additionally, we introduce an additional class center separation constraint loss to increase the distances between inter-class node representations, further enhancing their separability. Extensive experiments conducted on eleven graph benchmarks demonstrate that CGGSA achieves competitive or superior performance in node classification tasks, effectively alleviating over-smoothing and enhancing the performance of deep GNNs.