Graph Self-Supervised Learning: A Hybrid Approach Combining Contrastive and Generative Paradigms

Abstract

Graph-structured data is ubiquitous, yet labeled graph data remains scarce and expensive, limiting the effectiveness of supervised graph neural networks (GNNs). To address this, self-supervised learning (SSL) has emerged as a promising paradigm to pre-train GNNs on unlabeled graphs. However, existing SSL methods typically rely on a single pretext task, either contrastive or generative, leaving the potential of a synergistic combination underexplored. This paper conducts a systematic literature review and proposes a novel hybrid framework that integrates contrastive learning (GraphCL-style) with generative learning (GraphMAE-style). Specifically, the proposed method introduces three key innovations: curriculum masking, adaptive loss weighting, and Laplacian positional encoding, tailored to handle graph sparsity and class imbalance. Experiments on the Cora citation network demonstrate that the proposed hybrid model achieves 79.6% accuracy, closing 66% of the performance gap to fully supervised GCN without using any node labels. The findings indicate that hybrid SSL significantly reduces label dependency and offers a robust foundation for scalable graph foundation models.