Topology-Aware Self-Supervision for Node Embeddings with Sparse Labels and Noise

Abstract

The widespread application of graph-structured data in diverse domains such as biological networks, social media analysis, and financial fraud detection has driven significant advancements in graph representation learning. Graph Neural Networks have emerged as the dominant paradigm for learning node embeddings by aggregating structural and semantic information. However, the performance of traditional supervised Graph Neural Networks degrades sharply when confronted with real-world constraints, specifically the sparsity of high-quality node annotations and the pervasive presence of label noise. To address these critical limitations, this paper proposes a novel framework focused on topology-aware self-supervision for node embeddings under sparse and noisy conditions. By leveraging the intrinsic structural properties of the graph, our methodology constructs a robust self-supervised signal that guides the representation learning process without relying heavily on clean labels. We introduce a specialized topological encoder that preserves local neighborhood structures while applying adaptive data augmentations derived from node centrality metrics. Furthermore, we design a noise-resistant pseudo-labeling mechanism that synergizes with contrastive learning objectives to filter out corrupted supervisory signals. Through comprehensive theoretical analysis and extensive empirical evaluations across multiple benchmark datasets, we demonstrate that the proposed topology-aware mechanism significantly outperforms existing state-of-the-art baselines in scenarios characterized by minimal label availability and high noise ratios. The results indicate a robust capacity to capture underlying community structures, thereby offering a highly scalable and resilient solution for complex network analysis tasks.