This paper proposes a novel joint embedding prediction framework (JPEB-GSSL) to address computational inefficiency, contrastive target dependency, and representation collapse in graph self-supervised learning (SSL). We overcome the limitations of existing methods—feature reconstruction, speech sampling, and reliance on complex decoders—and present a non-contrast, view-invariant joint embedding prediction architecture that preserves semantic and structural information without contrastive targets or speech sampling. Furthermore, we introduce a semantic target term that integrates pseudo-labels derived using Gaussian Mixture Models (GMMs) to evaluate the contribution of latent features, thereby enhancing node discriminability. By leveraging single-context and multi-target relationships between subgraphs, we outperform existing state-of-the-art graph SSL methods across various benchmarks. This provides a computationally efficient and collapse-resistant paradigm for combining spatial and semantic graph features.
